Arista Networks
The switches that let thousands of AI chips talk to each other, every one running Arista's own software.
The scores
Growth3/4 ↑
40% growth, organic and unbought - but it narrowed into two customers, one of which went from 18% of sales to 26%.
Business quality3/4 →
The moat binds everyone below the largest two, who wrote their own operating systems - and an independent count still ranks Arista third.
Financial position4/4 →
$13.3 billion of cash, no debt of any kind, and operations funded $2.8 billion of half-year cash against $84.2 million of capital spending.
Execution record4/4 ↑
Eight consecutive beats and three upward revisions to one year's guide, with the supply problem disclosed before it bit rather than after.
Per share capture3/4 ↓
Diluted share count flat over a year on 4.2% stock compensation - but the buyback stopped, so issuance is no longer offset.
Potential3/4 →
$8.4 billion of performance obligations is committed and 91% lands within two years, but the 2030 prize needs share it does not hold.
Risk profile2/4 ↓
Two customers were 42% of sales, each running its own switch operating system for a decade, and one supplier is the sole chip source.
Valuation2/4 →
The price needs revenue to double from the guide by 2029 with the margin held, and still pays 26 times that result.
Evidence quality3/4 →
Central claims are filed or independently measured, but the guide, the AI goal and the margin band are management's and no filing will check them.
The Verdict
We rate Arista neutral. It sells network switches for AI data centres, all running one piece of its own software, and sales are growing fast, so the business is solid. The price assumes sales roughly double by 2029 with the margin held and Arista staying the leader, so it is full, not cheap. Yet an independent count ranks it third in AI switching, and its own filing says growth comes with bigger discounts to its largest customers. The view hinges on whether third place is timing or lost ground, which the margin and that ranking will show over the next few quarters.
The long caseAI data centres need far more network links, because every AI chip must talk to every other; Arista must keep its share of that spending as rivals gain.
The evidence under this - 8 claims
- The guide is a supply forecast rather than a demand forecast on the company's own account: the chief financial officer says 'Jayshree and I are guiding based on what we're confident we can get the supply for', which means it can move on a supplier's decision rather than a customer's
- The same guide was $10.5 billion at the analyst day, so the company's own visibility a year out was 17% below where it now sits - which cuts both ways but is not a record of precision
- The multiple is struck on trailing profit for a company guiding 40% growth; on the guided 2026 revenue of $12.6B at the guided 48-49% operating margin the same enterprise value is about 40 times the current year
- The research pack's own enterprise value of $257.8 billion nets only cash equivalents and ignores the $11.05 billion of marketable securities, so any enterprise multiple quoted from the pack is overstated - including the cohort comparison, where every member is computed the same way
- 2026-09-26editedre-struck at the 25 September 2026 close ($206.55, from $199.10) on the refreshed pack: enterprise value $246.7B, 54 times trailing operating profit rather than 52, and about 40 times the guided year rather than 39
- 2026-09-26editedre-read on the pack refreshed 26 September 2026: the 2027 consensus earnings per share moved from $5.22 to $5.23; revenue and the other years unchanged
- Dell'Oro qualifies the same ranking in the same sentence: Arista 'would have been more closely tied if deferred AI revenue were included'. Arista's deferred revenue balance is $6.87B and it says the majority is product related, so the gap between recognised revenue and shipped demand is large and real
- The ranking measures AI back-end switching only. Arista's front-end data centre, campus and routing revenue is the majority of the company and is not in it
- Gross margin rose sequentially, from 61.9% in the first quarter of 2026, so the trend within 2026 is up rather than down
- The chief financial officer attributes 20 to 30 basis points of that sequential improvement to a tariff refund rather than to the business, and says price increases will not reach the reported figure until late 2026 or 2027 because Arista is 'still going through a lot of backlog sales'
- 2026-09-26editedre-read on the pack refreshed 26 September 2026: the split is unchanged at 41 Buy, 12 Hold, none Sell
The company
The chips in an AI data centre swap results with each other constantly, through switches: boxes with dozens to hundreds of high-speed ports. Arista sells those boxes, and the same software, EOS, runs on every one, from an office cupboard to an AI cluster.
It designs none of the chips and builds none of the boxes. Broadcom supplies the silicon, with no written guarantee, and contractors assemble them.
The evidence under this - 4 claims
- 2026-09-27editedRe-sourced to the June 2026 10-Q, whose risk factors repeat the passage word for word; the fact now stands as of 30 June 2026 rather than the year end.
How it makes money
Arista is paid when it ships a switch, and again for multi-year support. Much of it must first pass the customer's acceptance test, so cash often arrives before the sale counts: $6.87B had been billed and not yet earned at the end of June. Two customers bought 42% of everything it sold in 2025.
The evidence under this - 3 claims
The chief financial officer says the balance 'can move significantly on a quarterly basis, independent of underlying business drivers', so it is not a backlog in the ordinary sense and a fall would not necessarily mean demand fell
- 2026-09-27editedReclassified from current state to historical: the three annual shares are closed periods that never move, and the June 2026 10-Q restates the same table with no interim figure, so it was never a statement about today.
Where the money comes from, stream by stream
| Stream | Share of sales (FY2025) | How it earns it |
|---|---|---|
| Core cloud, AI and data centre | about 65% | Switches and routers running EOS for data centres and AI clusters, sold mostly to the largest cloud and AI operators. |
| Campus and routing adjacencies | about 18% | Campus switches, Wi-Fi, VeloCloud branch networking and wide-area routers sold to enterprises and service providers. |
| Network software and services | about 17% | Support contracts and subscription software such as CloudVision, renewed yearly on the installed base. |
Who runs it
Jayshree Ullal has run Arista since 2008 and took it to $9 billion of sales; co-founder Ken Duda runs the technology, and two recent hires from Fastly and Aspen Technology run operations and finance. Insiders own under 3%, mostly Ullal's; a family trust outside management holds about 15%. Bonuses and performance shares pay on sales and operating profit for the year, none on the share price or on cash, and last year paid the maximum.
The evidence under this - 3 claims
Insiders sold 6,988,583 shares and bought none in the twelve months to 26 September 2026 (C51), so the stakes in the April table are falling.
- The 2025 maximum was set at $8.5 billion of revenue and $3.95 billion of non-GAAP operating income, both below what was delivered, so the ceiling was reached without the stretch the cap implies.
- Nothing in either plan pays on the share price, cash flow or return on capital, and only 62% of votes cast approved the pay programme at the 2025 meeting.
Why it wins, and who could beat it
Why customers come back
A customer does not adopt a switch; it adopts a way of running its network. Its automation, monitoring and staff training are written once for EOS, so leaving means rewriting them and retesting new hardware.
It solves a real problem. Training one AI model, Meta counted 35 of 419 unplanned stoppages in 54 days on switches and cables; one fault can restart the whole job.
The lock-in stops at the two biggest customers, who wrote their own switch software a decade ago and still buy Arista hardware to run it.
The evidence under this - 7 claims
- Microsoft has been one of Arista's two greater-than-10% customers in each of the last three years, and SONiC has existed since 2016 - so a decade of coexistence is the observed outcome rather than displacement
- Arista is not among the premier members the Linux Foundation lists (Alibaba, Broadcom, Dell, Google, Intel, Microsoft, NVIDIA), which is a statement about governance rather than about whether Arista's hardware runs it
- The same release positions migration back to EOS as the benefit - 'Arista platforms running SAI can be easily migrated to EOS, providing a flexible path to the mainstream, feature-rich EOS software' - so Arista treats SONiC as an on-ramp rather than a surrender
- Arista's 10-K never mentions SONiC or SAI by name and gives no revenue split, so how many ports ship on somebody else's operating system is not disclosed anywhere
Being chosen as the co-development partner on a customer's own switch programme is evidence of platform engineering the customer could not do alone, which argues the relationship is deeper than a hardware purchase rather than shallower
- 2026-09-27editedReclassified as historical: it describes the hardware Meta built in 2021, a closed fact, not the state of its network today.
- Revenue share says nothing about which software those ports run. Arista discloses no EOS-versus-SAI split, so a growing hardware business and a shrinking software attach would look identical from outside
- The two customers' spending has grown enormously over the same decade - Meta alone plans $130 billion to $145 billion of capital expenditure in 2026 - so a constant revenue share can coexist with a falling share of their switch ports
The same paper reports more than 90% effective training time across every interruption, so the whole cost of the network faults sits inside a loss under 10%; and it does not say which switch, Arista's or Meta's own, caused which fault.
The threat
Dell'Oro, an independent research firm, ranked Arista third in switches for AI networks in both of the last two quarters. First was Celestica, which sells unbranded boxes to the biggest buyers, including Meta. Second was NVIDIA, which sells its network with its chips, so the buyer never holds a switch contest. Cisco, fourth, gained the most ground.
The evidence under this - 5 claims
- Dell'Oro qualifies the same ranking in the same sentence: Arista 'would have been more closely tied if deferred AI revenue were included'. Arista's deferred revenue balance is $6.87B and it says the majority is product related, so the gap between recognised revenue and shipped demand is large and real
- The ranking measures AI back-end switching only. Arista's front-end data centre, campus and routing revenue is the majority of the company and is not in it
Being chosen as the co-development partner on a customer's own switch programme is evidence of platform engineering the customer could not do alone, which argues the relationship is deeper than a hardware purchase rather than shallower
- 2026-09-27editedReclassified as historical: it describes the hardware Meta built in 2021, a closed fact, not the state of its network today.
The rivals: CLS, NVDA, CSCO, HPE, MSFT, META, EXTR, DELL, KEYS, NTCT.
Who else does this, rival by rival
| Ticker | Rival | Where it competes | How far along | Priced against it |
|---|---|---|---|---|
| CLS | CelesticaPriced against it: Yes | AI Centers | On sale | Yes |
| NVDA | NVIDIA Spectrum-X, InfiniBand and NVLinkPriced against it: Shown, not counted | AI Centers | On sale | Shown, not counted |
| CSCO | CiscoPriced against it: Yes | Data Centers and WAN Centers | On sale | Yes |
| HPE | Hewlett Packard Enterprise, with the Juniper portfolioPriced against it: No usable figure | Campus Centers | On sale | No usable figure |
| MSFT | Microsoft, with SONiCPriced against it: Different product | Data Centers and WAN Centers | On sale | Different product |
| META | Meta, with FBOSS and its own OCP switch designsPriced against it: Different product | AI Centers | On sale | Different product |
| EXTR | Extreme NetworksPriced against it: Yes | - | On sale | Yes |
| DELL | Dell TechnologiesPriced against it: Shown, not counted | - | On sale | Shown, not counted |
| KEYS | Keysight TechnologiesPriced against it: One segment only | - | On sale | One segment only |
| NTCT | NetScout SystemsPriced against it: One segment only | - | On sale | One segment only |
How big it could get
How big this could get
Dell'Oro forecasts spending on switches for AI networks to pass $100B a year by 2030, and about two-thirds of AI switch sales today are the Ethernet kind Arista sells. That is a market of about $67B, against Arista's $10.54B of sales over the last four quarters.
The evidence under this - 3 claims
- A five-year forecast from a market research firm is a forecast, not a measurement, and this one is made at the top of a capital-spending cycle whose duration nobody has established
- The same firm expects co-packaged optics adoption 'spearheaded by NVIDIA', which would move value toward the party that also supplies the accelerators
InfiniBand sales more than tripled in the first quarter of 2026, which is the opposite direction from the one the forecast assumes, and Dell'Oro attributes a significant part of it to brownfield upgrades of an installed base that therefore still exists
Who has signed
Customers have committed $8.4B of future sales, 91% of it due within two years. But only $875M is described as binding. Most of the rest is the $6.87B already billed and waiting for customers to accept the equipment. Beyond that, nothing is ordered.
The evidence under this - 2 claims
The chief financial officer says the balance 'can move significantly on a quarterly basis, independent of underlying business drivers', so it is not a backlog in the ordinary sense and a fall would not necessarily mean demand fell
Where the growth is coming from, business by business
Where the growth is coming from
| Business | Share of revenue | Growth: last quarter | Last half-year | Last year | Share in 3 years at its half-year rate |
|---|---|---|---|---|---|
| Product | 86% | +38.8% | +37.7% | +33.2% | 88% |
| Service | 14% | +31.3% | +29.4% | +29.4% | 12% |
| Whole company | 100% | +37.7% | +36.5% | +32.6% | 100% |
On the filed split through 2026-06-30, if each business keeps its last half-year's growth, the company grows about 37% next year and 37% in year 3, against 32.6% over the last twelve months. A scenario, not a forecast: no business holds its rate.
The results
The numbers
Sales grew 33% over the twelve months to Jun 2026, about the same pace as the 29% in the year to Dec 2025.
Operating margin is 43%, up from 39% in the year to Dec 2023.
Running costs take 20¢ of each sales dollar, down from 32¢ in the year to Dec 2021.
Each dollar put into the business earns 29% a year after tax, up from 23% in the year to Dec 2021.
Yearly sales growth has ranged from +20% to +49% over the last 5 years.
The share count rose 0.8% in 3.5 years, and pay in shares costs 4.8% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-10-04. Fiscal years end in December.
| Dec 2021 | Dec 2022 | Dec 2023 | Dec 2024 | Dec 2025 | 12 mo to Jun 2026 | |
|---|---|---|---|---|---|---|
| Size and growth | ||||||
| Sales | $2.95B | $4.38B | $5.86B | $7.00B | $9.01B | $10.5B |
| Sales growth | +27% | +49% | +34% | +20% | +29% | +33% |
| Profit | ||||||
| Gross margin | 64% | 61% | 62% | 64% | 64% | 63% |
| Operating margin | 31% | 35% | 39% | 42% | 43% | 43% |
| Net margin | 29% | 31% | 36% | 41% | 39% | 38% |
| What growth costs, per sales dollar | ||||||
| Research and development | 20% | 17% | 15% | 14% | 14% | 13% |
| Sales and marketing | 9.7% | 7.5% | 6.8% | 6.1% | 5.9% | 5.5% |
| General and admin | 2.8% | 2.1% | 2.0% | 1.8% | 1.6% | 1.4% |
| Pay in shares | 6.3% | 5.3% | 5.1% | 5.1% | 4.9% | 4.8% |
| Capital spending | 2.2% | 1.0% | 0.6% | 0.5% | 1.3% | - |
| Cash | ||||||
| Cash from operations | $1.02B | $493M | $2.03B | $3.71B | $4.37B | - |
| Free cash flow | $951M | $448M | $2.00B | $3.68B | $4.25B | - |
| Balance sheet | ||||||
| Cash and short-term investments | $3.41B | $3.02B | $5.01B | $8.30B | $10.7B | $13.3B |
| Returns and shares | ||||||
| Return on capital | 23% | 29% | 32% | 30% | 28% | 29% |
| Shares (diluted) | - | 1.27B | 1.27B | 1.28B | 1.28B | 1.28B |
What growth costs
Growth has made Arista more profitable: research, selling and running costs rise more slowly than sales, so each new dollar of sales costs less to win. The squeeze is elsewhere, because selling more to the biggest buyers lowers what each box earns, so the gains show up below the gross margin, not in it.
The evidence under this - 2 claims
- Gross margin did not widen with them: it was 63.9% in 2020 and 64.1% in 2025 and fell to 62.9% in the second quarter of 2026 on sales to the largest customers (C32), so the leverage sits in operating costs only.
- The filed R&D line includes stock pay; management's own figure of about 11% of 2025 revenue is non-GAAP and excludes it, so the two definitions differ by roughly the stock pay charged to R&D.
- Gross margin rose sequentially, from 61.9% in the first quarter of 2026, so the trend within 2026 is up rather than down
- The chief financial officer attributes 20 to 30 basis points of that sequential improvement to a tariff refund rather than to the business, and says price increases will not reach the reported figure until late 2026 or 2027 because Arista is 'still going through a lot of backlog sales'
The last report (4 Aug)
Quarter ended June 2026, filed figures.
The second quarter of 2026 looked like another record: sales of $3.04B, up 37.7% on a year earlier, and a full-year forecast raised for the third time, to $12.6B.
The detail was in the margin. Gross margin, what is left after paying for the boxes, fell to 62.9% from 65.2%, and the filing says why: more sales went to the largest customers, who get bigger discounts. Other costs grew more slowly, so operating margin still rose. And the forecast now needs about 43% growth in each remaining quarter, faster than June's.
What it did to the thesisIt confirmed the growth is real and limited mainly by supply. It also gave the evidence behind the Neutral view: margins slipped because the largest customers get bigger discounts, and the raised forecast needs growth to speed up. Leadership and margin, which the price pays for, remain in question.
The quarter in detail
Sales growth from the previous quarter, filed figures. The company has restated its own history, so a quarter either side of the restatement may describe a different business; n/c marks one that does not compare.
Earnings per share $1.02 against $0.89 expected; $0.95 as filed, the gap being what the company adds back.
The evidence under this - 5 claims
- The guide is a supply forecast rather than a demand forecast on the company's own account: the chief financial officer says 'Jayshree and I are guiding based on what we're confident we can get the supply for', which means it can move on a supplier's decision rather than a customer's
- The same guide was $10.5 billion at the analyst day, so the company's own visibility a year out was 17% below where it now sits - which cuts both ways but is not a record of precision
- The acceleration is not a demand assumption but a supply one: management says second-half shipments depend on the supply-chain work done in the first half, and that if supply released further 'there is an opportunity to do better in the year'
- A rising deferred revenue balance means some of the second-half recognition is already billed and sitting on the balance sheet, so the required acceleration is less speculative than a bare order-book ramp would be
- Gross margin rose sequentially, from 61.9% in the first quarter of 2026, so the trend within 2026 is up rather than down
- The chief financial officer attributes 20 to 30 basis points of that sequential improvement to a tariff refund rather than to the business, and says price increases will not reach the reported figure until late 2026 or 2027 because Arista is 'still going through a lot of backlog sales'
- Operating leverage of this kind is a function of a revenue line growing 37.7%; it reverses when growth slows, and the expense base is people rather than a fixed asset, so it is a timing advantage rather than a structural one
- The measure is GAAP. On the non-GAAP basis management guides to, operating margin was 49.9% against 48.8% - a similar move, but struck after adding back $120.4 million of stock compensation in the quarter
What management did with the money
It holds $13.3B of cash and securities, no debt, and bought back no shares this year. Instead it has committed $9.7B to buying parts, nearly triple a year earlier, to secure supply in a market short of it.
Insiders in 2026Q2: 30 purchases and 178 sales, from their Form 4 filings.
The evidence under this - 4 claims
Stopping a buyback while the shares are at 64 times trailing earnings is a defensible allocation decision rather than a failure, and the company has no debt to service and no financing need that would force it
- 2026-09-26editedcounter-evidence re-struck at the 25 September 2026 close: trailing P/E is 64 on the refreshed pack, not 62
Arista's own risk factors say it expects inventory and purchase commitments 'to remain elevated and subject to volatility' and name 'a heightened risk of excess or obsolete inventory and supplier liability charges' - the company's own statement of what the commitment costs if demand moves
Arista's chief operating officer says memory supply is 'secured for 2026' with visibility 'well into 2027' across DDR4, DDR5 and NAND, which is a narrower and more favourable statement than the industry-wide one made two answers earlier on the same call
The price, and where we disagree
In short. Full. The price pays for Arista staying the market leader and holding its profit margin. The growth is real; the evidence questions both the leadership and the margin.
The share price does not make it urgent: Arista returned 44.4% over the past year, against 149.8% for the median company in its AI networking theme.
The test does. On 3 November Arista reports the September quarter, the first test of a forecast that needs growth to speed up. Underneath, spending on switches for AI networks passed the ordinary data-centre network for the first time in June, and both Arista and Dell'Oro say supply, not demand, limits it.
The evidence under this - 5 claims
- 2026-09-26editedre-struck on the pack refreshed 26 September 2026, taking both returns from the theme comparison's own window so the two figures are measured alike: 44.4% against a theme median of 149.8%, where the claim had set the price series' 32.1% beside the theme's 136.4% to 11 September
- The guide is a supply forecast rather than a demand forecast on the company's own account: the chief financial officer says 'Jayshree and I are guiding based on what we're confident we can get the supply for', which means it can move on a supplier's decision rather than a customer's
- The same guide was $10.5 billion at the analyst day, so the company's own visibility a year out was 17% below where it now sits - which cuts both ways but is not a record of precision
- The acceleration is not a demand assumption but a supply one: management says second-half shipments depend on the supply-chain work done in the first half, and that if supply released further 'there is an opportunity to do better in the year'
- A rising deferred revenue balance means some of the second-half recognition is already billed and sitting on the balance sheet, so the required acceleration is less speculative than a bare order-book ramp would be
- A five-year forecast from a market research firm is a forecast, not a measurement, and this one is made at the top of a capital-spending cycle whose duration nobody has established
- The same firm expects co-packaged optics adoption 'spearheaded by NVIDIA', which would move value toward the party that also supplies the accelerators
Arista's chief operating officer says memory supply is 'secured for 2026' with visibility 'well into 2027' across DDR4, DDR5 and NAND, which is a narrower and more favourable statement than the industry-wide one made two answers earlier on the same call
How we read the multiple
Buyers pay more than 50 times the operating profit Arista earned in the last year, the dearest of its rivals. Its biggest customers get bigger discounts, and gross margin fell from 65.2% to 62.9% in a year.
On analysts' own numbers, the price needs:
- Sales to roughly double, to about $24.1B by 2029.
- Operating profit held near 49 cents on each dollar of sales.
- Even then, today's price is about 26 times 2029's expected earnings: the margin is the bet.
Against its own history. Dearer for each dollar of sales than across its own recent years, on faster growth than then. Part of that is earned; the rest assumes the margin holds.
The evidence under this - 5 claims
- The multiple is struck on trailing profit for a company guiding 40% growth; on the guided 2026 revenue of $12.6B at the guided 48-49% operating margin the same enterprise value is about 40 times the current year
- The research pack's own enterprise value of $257.8 billion nets only cash equivalents and ignores the $11.05 billion of marketable securities, so any enterprise multiple quoted from the pack is overstated - including the cohort comparison, where every member is computed the same way
- 2026-09-26editedre-struck at the 25 September 2026 close ($206.55, from $199.10) on the refreshed pack: enterprise value $246.7B, 54 times trailing operating profit rather than 52, and about 40 times the guided year rather than 39
- 2026-09-26editedre-read on the pack refreshed 26 September 2026: the 2027 consensus earnings per share moved from $5.22 to $5.23; revenue and the other years unchanged
- Four contributors is a thin consensus three years out and the estimate is as likely to be stale as wrong; the 2027 figure, with 18 contributors, puts the shares at about 40 times, which is a firmer number on a nearer year
- A business with no debt, $13.3 billion of cash and a 45% operating margin can justify a premium to the market on quality alone, so the multiple is not evidence of mispricing by itself
- 2026-09-26editedre-struck at the 25 September 2026 close ($206.55): 26 times 2029 consensus earnings rather than 25, and about 40 times 2027 rather than 38
The 2025 outcome was 64.1% on a GAAP basis, so holding the range means accepting a fall of up to two points, and the first half of 2026 has already come in at 62.4%
- Gross margin rose sequentially, from 61.9% in the first quarter of 2026, so the trend within 2026 is up rather than down
- The chief financial officer attributes 20 to 30 basis points of that sequential improvement to a tariff refund rather than to the business, and says price increases will not reach the reported figure until late 2026 or 2027 because Arista is 'still going through a lot of backlog sales'
Priced against its rivals
The competitors its own annual report names, each confirmed against the rival's filing. Measured on company value for each $1 of operating profit, cheapest first:
| Ticker | Company | company value for each $1 of operating profit |
|---|---|---|
| CSCO | Cisco Systems | 28.8× |
| CLS | Celestica | 30.6× |
| EXTR | Extreme Networks | 44.2× |
| ANET | Arista Networks | 56.7× |
| DELL | Dell Technologies not in the median | 27.5× |
| NVDA | NVIDIA not in the median | 27.7× |
| HPE | Hewlett Packard Enterprise Company held out | not priced |
The middle of the other 3 is 30.6. Dell Technologies Inc., NVIDIA Corporation compete with it in one product line inside a much bigger company, so their price says little about this business and they are not counted in the middle. Arista Networks, Inc. is dearest of 4.
Rivals' figures come from a data provider, so read the spread of the group rather than any single pair.
The dearest of its filed rivals for each dollar of operating profit, by a wide margin. That standing is Arista's own, not the group's.
Private or foreign rivals that cannot be priced: Huawei Technologies Co., Ltd., Juniper Networks, Inc., Darktrace, ExtraHop, Gigamon.
Where we disagree with the market
Arista is the pure-play winner of AI networking, and the price is what that leadership costs.
The growth is real and limited only by supply. The leadership the price pays for is not what the independent count shows.
Most analysts rate the shares a buy and none a sell. But two public facts sit outside that story: Dell'Oro's ranking, and the filing's own reason the margin is falling.
The market has doubted before, selling three strong reports in fifteen months over concentration, margin and supply. Sales outgrew all three; only the margin worry came true.
The evidence against us is real. Dell'Oro says Arista would rank higher counting its billed-but-unearned sales, and gross margin holding at 64% while sales grow above 30% would bound the discounts.
What analysts say, and the evidence
The evidence under this - 7 claims
- 2026-09-26editedre-read on the pack refreshed 26 September 2026: the split is unchanged at 41 Buy, 12 Hold, none Sell
- Dell'Oro qualifies the same ranking in the same sentence: Arista 'would have been more closely tied if deferred AI revenue were included'. Arista's deferred revenue balance is $6.87B and it says the majority is product related, so the gap between recognised revenue and shipped demand is large and real
- The ranking measures AI back-end switching only. Arista's front-end data centre, campus and routing revenue is the majority of the company and is not in it
- Gross margin rose sequentially, from 61.9% in the first quarter of 2026, so the trend within 2026 is up rather than down
- The chief financial officer attributes 20 to 30 basis points of that sequential improvement to a tariff refund rather than to the business, and says price increases will not reach the reported figure until late 2026 or 2027 because Arista is 'still going through a lot of backlog sales'
The chief financial officer says the balance 'can move significantly on a quarterly basis, independent of underlying business drivers', so it is not a backlog in the ordinary sense and a fall would not necessarily mean demand fell
The twenty sessions after February 2025 ran into a broad technology sell-off, and the May 2026 fall followed a 34% run in the month before, so neither is all objection; a 7.1% fall after the November 2024 print is left out because no stated objection was found for it.
Two accounts of 5 November 2025 disagree on the cause, one naming the margin guide and one the growth guide, so what the market objected to that day is what coverage said, not a measured fact.
What could go wrong, and how we would know
The risks
- Third, not first. Dell'Oro ranks Arista behind Celestica and NVIDIA in AI switching for two quarters running, with Cisco gaining the most.
- Two customers, 42% of sales. One was 26% of 2025 sales, up from 18% in 2023. The chief executive expects only one, maybe two, above 10% this year.
- Insiders only sell. Insiders sold 521 times in the past year, the chief executive and the co-founder among them, and never once bought on the open market.
The evidence under this - 6 claims
- Dell'Oro qualifies the same ranking in the same sentence: Arista 'would have been more closely tied if deferred AI revenue were included'. Arista's deferred revenue balance is $6.87B and it says the majority is product related, so the gap between recognised revenue and shipped demand is large and real
- The ranking measures AI back-end switching only. Arista's front-end data centre, campus and routing revenue is the majority of the company and is not in it
- Gross margin rose sequentially, from 61.9% in the first quarter of 2026, so the trend within 2026 is up rather than down
- The chief financial officer attributes 20 to 30 basis points of that sequential improvement to a tariff refund rather than to the business, and says price increases will not reach the reported figure until late 2026 or 2027 because Arista is 'still going through a lot of backlog sales'
- 2026-09-27editedReclassified from current state to historical: the three annual shares are closed periods that never move, and the June 2026 10-Q restates the same table with no interim figure, so it was never a statement about today.
The same answer says Arista remains 'very committed' to both and that she wants flexibility on 'who are the 10% customers as we ship', so the sentence is as consistent with a third customer arriving as with one of the two leaving
Only one option exercise appears among the 536 rows, so these are sales of shares already held rather than same-day exercises - but long-tenured founders and officers of a company whose shares have risen sell under pre-arranged plans as a matter of course, and the transaction codes cannot distinguish a plan sale from a discretionary one
- 2026-09-26editedre-struck on the pack refreshed 26 September 2026 (521 sales and 6,988,583 shares, from 517 and 6,759,250), and corrected: the claim said every transaction was a sale, but the window also holds transfers and an option exercise, and the sellers are four outside directors plus the chief financial officer rather than 'five directors'
What we still do not know
- Whether third place is lost ground or equipment still awaiting acceptance. Nobody publishes the figure that would settle it.
- How much of the $6.87B billed but not yet earned is AI equipment. The company says only that most of it is product.
- What the AI business earns. Arista reports as one segment, so no quarterly AI figure is filed.
The evidence under this - 3 claims
- Dell'Oro qualifies the same ranking in the same sentence: Arista 'would have been more closely tied if deferred AI revenue were included'. Arista's deferred revenue balance is $6.87B and it says the majority is product related, so the gap between recognised revenue and shipped demand is large and real
- The ranking measures AI back-end switching only. Arista's front-end data centre, campus and routing revenue is the majority of the company and is not in it
The chief financial officer says the balance 'can move significantly on a quarterly basis, independent of underlying business drivers', so it is not a backlog in the ordinary sense and a fall would not necessarily mean demand fell
What would change our view
- Neutral→AttractiveThe demand holds and the multiple comes in
It needs all of:
- ○Sales growth stays above 20% with operating margin at 45% or better, and billed-but-unearned sales keep rising.checked at the next results, due by 3 Nov
- ○The price falls to about 30 times this year's forecast operating profit, so sales no longer need to double by 2029.about ≈$153.38 today, 26% below the last close
- Neutral→UnattractiveConcentration or ranking turns
It needs any of:
- ○Either of the two largest customers falls below 10% of sales because it bought less.checked at the next results, due by 3 Nov
- ○Dell'Oro ranks Arista fourth or lower, or gross margin stays below 62% for two quarters running on customer mix.checked at the next results, due by 3 Nov
We would also be wrong if
- Billed-but-unearned sales fall for two quarters running while growth slows.
- Sales growth settles below 20% while operating margin falls under 45%.
What is coming
- 3 November 2026 Q3 2026 results: Gross margin against 62.5% and revenue against the guided $3.3 billion, with deferred revenue beside them, and whether the buyback resumes.
Every dated event ahead
- 3 November 2026 Q3 2026 results: Gross margin against 62.5% and revenue against the guided $3.3 billion, with deferred revenue beside them, and whether the buyback resumes.
- December 2026 1.6 terabit switch trials with a few very large customers: Whether Arista stays a generation ahead at the biggest buyers; production is not expected until 2027, so a slip would show in trials before sales.
- February 2027 Q4 2026 results and the 2026 10-K: Whether the $12.6 billion guide was delivered, how many customers top 10% and how large the biggest grew, and the 2027 margin guide.
- April 2027 2027 proxy statement: What the 2026 bonus and performance shares paid against their revenue and operating-income goals, and whether the 62% say-on-pay vote changed the design.
Since we wrote this
- 8 OctNewsArista launches Ethernet racks for AI scale-up, with partners named and no customers. View heldArista announced rack-scale Ethernet architectures for scale-up and scale-out AI.
- 5 OctThesis reviewThe long case now sits under the verdictAfter the five-name pilot, the owner asked for the long case on every thesis.
- 4 OctThesis reviewOwner's checklist: revenue streams, cost of growth, management, catalystsThe owner's checklist found the method silent on who runs the company, what growth costs, where the money comes from and what is dated ahead; this pass fills those four blocks from the proxy and filings.
- 27 SepThesis reviewThree past sell-offs and a buyer's own number added; view heldNothing new happened. The review filled in the record: the market sold three beats on three different worries, and only the margin worry came true.
- 26 SepThesis reviewPrice re-struck at $206.55; two data slips fixed, view heldA data fix refreshed the numbers. The shares cost 54 times trailing operating profit rather than 52, revenue beat in seven of eight prints, not eight, and nothing the view rests on moved.
- 11 SepThesis writtenNeutral, medium convictionAn independent count ranks Arista third in AI switching, while the price pays for a leader holding its margin.
- 11 SepThesis reviewLock-in found not to bind the two biggest customers. View heldBoth write their own switch software, so the lock-in does not hold 42% of sales. But both kept buying for a decade, and the price case is untouched.
Also checked 1 time with nothing new to act on.
Sources 29
Sources - 29
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-02-17Item 1 in full - the product and EOS description, the Competition section naming Cisco, Dell/EMC, Extreme, HPE, Huawei, Juniper, Nvidia and white box vendors, the Broadcom sole-silicon dependency, the three contract manufacturers, the customer concentration table and the qualification-cycle languagebacksC01C02C03C04C05C12C26C59C64C67C71
- S02primarycompany filing2026-08-05the statements, the balance sheet, the deferred revenue and remaining performance obligation note, the $9.7B purchase commitment note, the buyback disclosure, the single-segment note, and the MD&A sentence attributing the gross margin fall to large-customer discountsbacksC02C05C21C22C23C24C25C26C28C32C36C38C40C41C42C43C44C48C53C59C64C66C67
- S03primarycompany filing2026-08-04the GAAP to non-GAAP reconciliation, the gross and operating margin series against the prior year, and the third-quarter guidancebacksC08C28C32C43C53
- S04primarytranscript2026-08-04Arista second quarter 2026 earnings call transcriptthe $12.6B full-year guide and its two prior versions, the AI and campus goals, the purchase-commitment history, the supply-chain answers, and fourteen analyst questions including the white box, scale-up and visibility exchanges
Given our improving stance in supply chain, we are excited to increase our guidance for the third time this year to $12.6 billion revenue in 2026. We are now projecting 40% annual growth, which is an incremental $2.1 billion over our Analyst Day goal of $10.5 billion and an incremental $1.1 billion over our recent projections of $11.5 billion in May of 2026. ... Arista is leaning in with our increased multiyear purchase commitments, now almost tripling from a year ago at $3.6 billion to approximately $9.7 billion by the end of Q2 2026. ... we still have the same kind of 2 quarters of visibility that we referred to. But what we would say is given where we are at this time of year, Jayshree and I are guiding based on what we're confident we can get the supply for. ... The industry is going to have a 2-year problem. I don't think we get out of it as an industry until 2028. ... we remain very committed to our 2 longest partners, Microsoft and Meta. And I fully expect there to be 1, maybe 2 10% customers. ... where NVIDIA provides the full vertical stack and usually, that's an NVLink. So there's very little participation from Arista or anybody else's scale up. ... Our AI fabrics momentum with Etherlink switches now exceeds 100 cumulative customers from the initial 4 to 5 customers I spoke of in 2024. ... On the 1.6T, if I -- I'll go back, history is always a great lesson for what will happen. If I look at the 800 T, we were in trials in '23, '24 and the real ramp came to us in '25 and '26. So it takes about 6 months, maybe even a year sometimes. ... So we will be in trials in the second half of this year. In terms of customers, it will be single digits, but very large customers. And I think the real production will still be in 2027. ... our 2026 campus revenue goal is at least $1.25 billion and our AI fabrics goal is at least $3.5 billion. For gross margin, we are maintaining the range for the fiscal year of 62% to 64%, inclusive of mix and anticipated supply chain cost increases for memory and silicon. ... Our total deferred revenue balance was approximately $6.9 billion, up from $6.2 billion in the prior quarter. The majority of the deferred revenue balance is product related. ... In the quarter, we did not repurchase our common stock. Of the $1.5 billion repurchase program approved in May 2025, $817.9 million remain available for repurchase in future quarters.
backsC06C07C14C23C27C30C33C36C39C42C54C64 - S05primarycompany filing2026-02-12the full year 2025 figures, the Gartner placements, the VeloCloud acquisition, the executive appointments and the first-quarter 2026 guidancebacksC08C21C29C46C48
- S06primaryindustry data2026-09-03Dell'Oro Group - AI Back-End Networks Switch Sales Surpass Front-End Networks for the First Time in 2Q 2026 Dell'Oro Groupthe vendor ranking on the quarter Arista has just reported, the front-end/back-end crossover, and the supply-constrained call
REDWOOD CITY, Calif. - September 03, 2026 - ... 2Q 2026 marked a major inflection point as switch sales in AI back-end networks surpassed those in front-end networks for the first time. ... "AI back-end networks have grown at such a rapid pace that, within just three years of their emergence, spending surpassed that of the large, well-established front-end network market" ... "We expect the market to remain supply-constrained, rather than demand-constrained, for at least the next one-to-two years." ... Celestica retained the leading position in Ethernet AI back-end switch sales, followed closely by NVIDIA. Arista ranked third, but would have been more closely tied if deferred AI revenue were included. Cisco gained the most share and ranked fourth. ... 800 Gbps switches accounted for the vast majority of the Ethernet switch shipments and revenues in AI back-end networks during the quarter. 1.6 Tbps began sampling and is expected to ramp in the second half of 2026.
backsC09C18C39C58 - S07primaryindustry data2026-06-02Dell'Oro Group - Ethernet Extends Lead in AI Scale-Out Networks Despite Strong InfiniBand Rebound Dell'Oro Groupthe same ranking a quarter earlier, the Ethernet share of AI-cluster switch sales, and the InfiniBand rebound
REDWOOD CITY, Calif. - June 02, 2026 - ... Ethernet switch sales in AI back-end networks more than doubled and accounted for about two-thirds of data center switch sales in AI clusters during the first quarter 2026. ... "InfiniBand sales more than tripled during the quarter, supported by the ramp of 800 Gbps switches shipping with NVIDIA's Blackwell Ultra platform. However, while greenfield deployments may have contributed to some of this growth, we believe a significant portion of this demand was tied to upgrades of the installed base in brownfield deployments rather than greenfield expansion." ... Celestica regained the leading position during the quarter, followed closely by NVIDIA. Arista ranked third, despite a significant portion of its AI-related product revenue being deferred. Including Arista's deferred revenue for the quarter, its AI back-end network sales would have been very close to those of Celestica and NVIDIA. Cisco recorded the largest share gain during the quarter and ranked fourth.
backsC10C13C19 - S08primaryindustry data2026-02-04the size of the prize measured from outside the company, the port-speed migration, and the scale-up read on NVLink, UALink and Ethernet
REDWOOD CITY, Calif. - February 04, 2026 - ... spending on data center switches deployed in AI back-end networks is forecast to surpass $100 billion by 2030, driven by expanding deployments across scale-up, scale-out, and scale-across domains. While multiple technologies are expected to coexist and see strong adoption over the forecast period, Ethernet is projected to dominate both the scale-up and scale-out segments of the market. ... "Scale-up compute fabrics have been historically dominated by proprietary fabrics such as NVLink but just as Ethernet prevailed over InfiniBand in large-scale scale-out environments, we are now seeing alternative technologies such as UALink and Ethernet gain momentum in scale-up compute fabrics." ... The majority of switch ports deployed in AI back-end networks have shifted to 800 Gbps, and are expected to be 1600 Gbps by 2027 and 3200 Gbps by 2030. ... Neo Cloud will be the fastest growing customer segment over the next five years.
backsC18C19 - S09primarycompany filing2026-09-02a rival naming Arista back, its own AI-infrastructure positioning, and the $34,668M of networking product revenue that sets the perimeter for the cohortbacksC17
- S10primarycompany filing2026-08-17a rival's own ranking of the switching market, which puts Cisco, HPE and Huawei above Arista and names Arista only 'to a lesser extent'backsC16
- S11primarycompany filing2026-08-26the single Compute & Networking segment, and the absence of any separate networking revenue line - the disclosure that would size this competitor no longer existsbacksC15
- S12primarycompany filing2026-08-26the scale of the memory pre-buy by the largest buyer in the market Arista is also buying memory in
Our commitments increased from $119 billion last quarter to $279 billion, primarily related to the procurement of memory.
backsC15C38 - S13primarycompany filing2026-02-27the white box switch business described by the company Dell'Oro ranks first in Ethernet AI back-end switching, and the CCS segment revenue that sets its perimeterbacksC11
- S14primarycompany filing2026-07-30a named Arista customer's own filed capital plan for 2026 - the demand this thesis depends on, stated by the party that will spend itbacksC20
- S15primarycompany filing2026-09-04the Networking segment revenue and its campus, routing and data-centre split, which sets the perimeter for the cohort after the Juniper acquisitionread, but no claim rests on it
- S16primaryvendor data2026-09-26Research pack, frozen at content/ANET/research-pack.json (refreshed 2026-09-26)the market tier: price, market value, the 252-day return and the theme comparison, the analyst split and consensus estimates, the insider filings, FINRA short interest, the volatility bucket, and the filed competitor cohort's multiplesbacksC46C48C49C50C51C52C53
- S17primaryindustry data2022-04-14Linux Foundation - Software for Open Networking in the Cloud (SONiC) Moves to the Linux Foundation Linux Foundationwho created the competing operating system, how widely it runs, and the premier member list - which does not include Arista
Created by Microsoft for its Azure data centers, SONiC is an open source network operating system (NOS) based on Linux that runs on over 100 different switches from multiple vendors and ASICs. ... "Microsoft founded SONiC to bring high reliability and fast innovation to the routers in Azure cloud data centers. We created it as open source so the entire networking ecosystem would grow stronger. SONiC already runs on millions of ports in the networks of cloud scalers, enterprises, and fintechs," said Dave Maltz, Technical Fellow and Corporate Vice President, Microsoft Azure Networking. ... SONiC brings a strong existing ecosystem, with premier members including Alibaba, Broadcom, Dell, Google, Intel, Microsoft, NVIDIA and 50+ global partners.
backsC55 - S18primarycompany release2017-03-08Microsoft Azure blog - SONiC: The networking switch software that powers the Microsoft Global Cloud MSFTthe customer's own account of why it wrote its own switch software, in its own words
Running one of the largest clouds in the world, Microsoft has gained a lot of insight into building and managing a global, high performance, highly available, and secure network. Experience has taught us that with hundreds of datacenters and tens of thousands of switches, we needed to: Use best-of-breed switching hardware for the various tiers of the network. Deploy new features without impacting end users.
backsC55 - S19primarycompany release2020-05-05Arista's own answer to the substitute: sell the hardware into it, and treat migration back to EOS as the upside
Arista Networks (NYSE:ANET) today strengthened its commitment to open network software with the introduction of Arista switches Powered by SONiC (Software for Open Networking in the Cloud). ... Distribution of Arista switches Powered by SONiC will be available as a network operating system in addition to Arista's standard EOS offering. Customers have the option of choosing between Arista EOS and Arista SAI with SONiC software on these systems, and as an added benefit, Arista platforms running SAI can be easily migrated to EOS, providing a flexible path to the mainstream, feature-rich EOS software for broader network roles. ... The open source SONiC operating system is already available and in production on Arista switching platforms for select customers.
backsC56 - S20primarycompany release2015-03-10the second of Arista's two largest customers building its own switch and its own switch operating system, eleven years ago
Last summer, we shared that the networking team at Facebook was developing its own data center switches and switch software. The switches and software are now in production across multiple data centers. Today, we are happy to announce the initial release of our Facebook open switching system (code-named "FBOSS") project on GitHub and our proposed contribution of the specification for our top-of-rack switch (code-named "Wedge") to the OCP networking project.
backsC57 - S21primarycompany release2021-11-09the customer co-developing an Arista switch alongside its own, and naming Celestica as the manufacturer of its in-house designs
Meta has developed two next-generation 200G fabric switches, the Minipack2 (the latest version of Minipack, Meta's own modular network switch) and the Arista 7388X5, in partnership with Arista Networks. ... The Arista 7388X5 is also based on the Broadcom Tomahawk4 25.6T switch ASIC, with versions of the 7388X5 also utilizing a Credo chipset. ... Both the Wedge400 and 400C are manufactured by Celestica and are open platforms that developers of any size, from startups to large ISPs, can utilize for their own projects. ... In the past, FBOSS, Meta's own network operating system for controlling network switches, has utilized the specific API provided by the ASIC vendor. Now, with FBOSS being adapted to OCP SAI and deployed at scale in the Meta network, we can work with more silicon vendors.
backsC58 - S22primaryother2024-07-31a buyer's own count of what network faults cost it: 35 of 419 unexpected interruptions in 54 days of training came from switches and cables, on a fabric that names the Arista 7800
Llama 3 405B is trained on up to 16K H100 GPUs, each running at 700W TDP with 80GB HBM3 ... Llama 3 405B used RDMA over Converged Ethernet (RoCE) fabric based on the Arista 7800 and Minipack2 Open Compute Project OCP rack switches. ... The synchronous nature of training makes it less fault-tolerant-a single GPU failure may require a restart of the entire job. ... During a 54-day snapshot period of pre-training, we experienced a total of 466 job interruptions. Of these, 47 were planned interruptions due to automated maintenance operations such as firmware upgrades or operator-initiated operations like configuration or dataset updates. The remaining 419 were unexpected interruptions ... Network Switch/Cable | Network | 35 | 8.4% ... Despite these challenges, for Llama 3, we achieved higher than 90% effective training time while supporting automated cluster maintenance
backsC61 - S23secondarynews2025-02-19the objection stated on the day of the February 2025 fall: the Meta account shrinking
Meta Platforms is one of Arista's largest, responsible for roughly 15% of the company's revenue. However, that figure is down from 20% last year.
backsC63 - S24secondarynews2025-11-05the objection stated on the day of the November 2025 fall: the fourth-quarter gross margin guide
Q4 revenue: $2.3-$2.4 billion (midpoint $2.35B vs. $2.33B expected). Gross margin: 62-63%, slightly below the previous quarter. ... a softer margin outlook, in-line revenue guidance, and a broader tech-sector pullback prompted short-term profit-taking
backsC63 - S25secondarynews2025-11-05a second, independent account of the same day naming the growth guide rather than the margin
Arista expects to generate between $2.3 billion and $2.4 billion in revenue in the fourth quarter, which represents year-over-year growth of about 22% at the midpoint. This expected slowdown could be what's driving the stock lower
backsC63 - S26secondarynews2026-05-05the objection stated on the day of the May 2026 fall: the multiple and the supply constraint
The stock is currently selling for 62 times earnings and 39 times next year's expected earnings, a lofty multiple to be sure. ... Add to that the company's capacity constraints, and fair-weather investors are engaging in a bout of profit-taking.
backsC63 - S27primaryvendor data2026-09-27FMP historical end-of-day prices for ANET, September 2023 to 25 September 2026the close before and after each of twelve results releases, and the change twenty and sixty sessions onbacksC62
- S28primarytranscript2026-02-12Arista fourth quarter 2025 earnings call transcriptthe 2025 revenue split by customer sector and by product line, the 2026 AI and campus goals as first set, and the 2025 buyback total and average price
Shifting to annual customer sector revenue for 2025, cloud and AI titans contributed significantly at 48%. Enterprise and financials recorded at 32%, while AI and specialty providers, which now includes Apple, Oracle, and their initiatives, as well as emerging Neo Cloud, performed strongly at 20%. ... Arista's core sector revenue was driven at 65% of revenue. ... The combined campus and routing adjacencies together contribute approximately 18% of revenue. Our third and final category is the network software and services based on subscription models, such as A-Care, CloudVision, observability, advanced security, and even some branch edge services. ... Arista's subscription-based network services and software revenue contributed approximately 17%, and please note that it does not include perpetual software licenses that are otherwise included in core or adjacent markets. ... Within fiscal 2025, we repurchased $1,600,000,000 of our common stock at an average price of $100.63 per share.
backsC65 - S29primarycompany filing2026-04-16beneficial ownership as of April 2, 2026, the executive officers and director biographies, the 2025 Bonus Plan metrics, weights and targets, the PRSU goals and 2025 achievement, the CEO's outstanding awards and the 2025 say-on-pay resultbacksC68C69C70