The Zhitong Finance App learned that in the early morning of November 4, Beijing time (after market on November 3 EST), cloud network equipment company Arista Networks (ANET.US) announced the results for the third quarter of 2026 ending September 30, and held a conference call after the market on the same day.
Third-quarter revenue was US$3.66 billion, up 20.5% month-on-month and 58.5% year-on-year, the highest quarterly year-on-year growth rate since the company went public. It exceeded the company's guideline of about US$3.3 billion, and also higher than Bloomberg's agreed estimate of US$3.33 billion. The non-GAAP gross profit margin was 64.0%, down 1.2 percentage points year on year and 0.6 percentage points higher than the company's previous guideline of about 63%; non-GAAP operating margin was 49.1% (48.6% in the same period last year); non-GAAP net profit of US$1.53 billion, up 56% year over year; non-GAAP diluted EPS of US$1.20, up 55.8% year over year, higher than Bloomberg's agreed expectation of $1.05. GAAP diluted EPS was $1.22, up 82.1% year over year, higher than the non-GAAP scale — according to Zhitong Finance's disassembly based on financial reports, mainly because GAAP results included about US$134.5 million in strategic investment income (non-GAAP excluded). Additionally, the company was included in the S&P 100 index in September.
In terms of stock prices, after the results were announced, the stock surged after the market, rising 12% at one point; as of press release, it had risen nearly 7% in night trading.
Market Focus and Management Response
This quarter's focus is not on Q3 results itself, but rather on the 2027 guidance first given by management: revenue of about US$18.5 billion (increase of 38% — 40%), which is significantly higher than Bloomberg's agreed estimate of US$16.4 billion; of these, the AI fabrics (AI fabrics) target is about US$7 billion, double the 2026 baseline of US$3.5 billion, and the cognitive campus (cognitive campus) target is about US$2 billion, an increase of 60% (2026 baseline of US$1.25 billion); gross margin guidance of 62% — 64%, operating margin of about 49%. Meanwhile, the company raised its 2026 full-year revenue guidance for the fourth time in the year to US$13.3 billion (+48% YoY).
CEO Jayshree Ullal characterized this round of demand as an AI supercycle: “In 2026, we are in the midst of an AI supercycle, and there are already more than 100 Ethernet AI customers. By the end of 2027, we anticipate more than 10 gigawatts of deployment, spanning more than 10 different types of compute accelerators. What a significant inflection point in just three years.” She recalled that when the company started three years ago, there were only 4-5 AI customers, about 50,000-100,000 NVIDIA GPUs, and the network was mostly bound to the InfiniBand proprietary system.
Regarding supply assurance concerns (Susquehanna's Mehdi Hosseini asks if all chips and storage have been locked down), President and COO Todd Nightingale responded: The company “has invested in large-scale long-term strategic agreements in silicon wafers, PCB production capacity, and memory” and is “very confident” that supply will be delivered according to the 2027 guidelines, but “we will not place all orders 5 quarters in advance for the whole of next year; not every order has been placed”. CFO Chantelle Breithaupt added that the company has visibility into demand for the next two quarters, and the increase in procurement commitments to US$11.4 billion is a sign of confidence.
The first 2027 guidelines were given: AI doubled to 7 billion US dollars, and the park increased by 60%
Management revealed the 2027 framework for the first time: revenue of around US$18.5 billion (+38% — 40%), the AI network business doubled to about US$7 billion, the fastest growing sector; the park business was about US$2 billion (+60%), “far faster than the market”; gross profit margin of 62% to 64%, operating margin of about 49%, and tax rate of about 21%. Corresponding baseline for 2026: Annual revenue of US$13.3 billion, of which approximately US$3.5 billion for AI and US$1.25 billion for the park.
Ullal explained why the overall growth rate in 2027 was lower than the AI sector's growth rate: “When some businesses grow 60% or even 100%, while overall growth is 38% to 40%, it's clear that some businesses aren't that fast. We are confident that all of our businesses will continue to grow in double digits.” She emphasized that this was not conservative: “Our starting guide was far higher than our previous predictions, enjoy it and hope the future will be better.” The company also revised the 2030 Serviceable Market (TAM) judgment to “more than 200 billion US dollars.”
AI network: 100+ customers, 10 GW deployment and scale-up schedule
According to Ullal, the company supports three types of AI networks: scale-up (in-rack), scale-out (in-cluster), and scale-across (across data centers). Scale-Out and Scale-Across are still the main revenue earners. This year, the two account for roughly “73 sales,” and in 2027, Scale-Across's share is “at least one-third, probably higher”. The main product is 7800 AI Spine. In terms of scaling up, the company showcased the rack-level scale-up design at the Open Computing Project (OCP) conference in October. It collaborated with partners such as AMD, Arm, Broadcom, D-Matrix, Meta, Microsoft, and Qualcomm to launch an EtherLink SU-144 architecture that supports 144 accelerators in a single rack and can be expanded to 1,024 xPUs across racks based on the Ethernet Scaling Network Standard (ESUN); scaling tests will begin in 2027, and mass production in 2028. It will be raised to 144 ports, and approximately 1,000 XPUs will be connected across racks. She anticipates that the 2027 scale-up revenue contribution “does not constitute a meaningful figure” and that contributions will begin in 2028.
Regarding the 10 gigawatt caliber (Goldman Sachs Michael Ng asked whether it was the number of industries or companies), Ullal responded: “Our guess is that by the end of 2027 — there will be an exponential climb in 2027 — more than 10 gigawatts will be deployed”, which includes both additional gigawatts of deployment and the use of existing data center gigawatts in the scale-across model; the company has now reached 4-5 types of xPUs. As scale-up progresses, about 5 new types will be added in 2027.
Enterprise and campus: security-driven, multi-year cycle
President and COO Todd Nightingale introduced the season's three representative winners: an Asian network services company chose Arista (100G/400G platform based on EVPN-VXLAN, which will be the front-end network for its AI deployment) due to the AI era's strategy for speed, scale, and resilience; a leading hospital group selected Arista after a strict zero trust assessment to achieve dynamic security segmentation with MSS (Macro Segmentation Service); a regional resort and casino used Arista routes to exchange plus campus access switches Update their mission-critical networks.
Founder and CTO Ken Duda placed security at the center of the AI narrative: the company collaborated with Anthropic, Google, OpenAI and others to embed cutting-edge model AI security capabilities into the vulnerability discovery process (such as “Glasswing” projects, the original transcript was vague, see checklist); EOS's single operating system architecture made its CVE vulnerability count “less than 100, which is at the lowest level in the industry for 20 years”, and Ullal said it was “10 to 50 times better than alternatives”. The reason for the high growth of the park, Duda boils down to “the enterprise market is coming to us” — almost every important enterprise regards the park and branches as a mission-critical network; Nightingale added that the park's equipment update cycle is very long, and it takes time for customers to recognize and switch. “This is expected to be a multi-year cycle”.
Financial details: Significant increase in R&D, stocking and promised expansion
CFO Chantelle Breithaupt revealed that international revenue was US$682.5 million, accounting for 18.6% (23% in the previous quarter). The month-on-month change was mainly due to the geographical structure of sales by major customers. The three regions all achieved double-digit year-on-year growth; operating expenses were US$545.3 million (accounting for 14.9% of revenue), of which R&D expenses amounted to US$408.5 million, accounting for 11.2% of revenue (US$278.1 million last quarter), and increased investment in liquid cooling, high power density switching and AI software - to Wolfe Research on R & When asked by amp; D jumped up, Duda explained that due to the demand for large-scale AI customer cooperation and infrastructure upgrades brought about by the introduction of liquid cooling technology, Breithaupt maintains a framework where R&D accounts for 9% to 11% of revenue; sales and marketing expenses account for 3.0%.
Operating cash flow was approximately US$1.2 billion; inventory was US$3.2 billion (2.5 billion in the previous quarter); procurement commitments were US$11.4 billion (prior quarter), including newly signed multi-year memory agreements; deferred revenue was approximately US$6.8 billion (6.9 billion in the previous quarter), and deferred product revenue decreased by US$86 million month-on-month. In response to UBS's question about the relationship between deferred revenue and 38% — 40% guidance, Breithaupt responded: “Deferred revenue had the largest inflow and outflow rate this quarter, but the balance remained flat, while achieving 58.5% revenue growth, which is a sign that demand is very healthy”; she also expects the balance of deferred revenue at the end of the year to be higher than the end of last year. In response to Bank of America's sharp question about “decoupling orders and deferred revenue,” Ullal said bluntly: “You want to disassemble demand into a physical problem, and I have no answers... What matters is the general direction: we all have needs for AI, enterprises, campuses, and data centers. This was the best time in my 18-year career for the company.”
Regarding BNP Paribas's question on whether the 2027 guidelines rely on building new power facilities or using deferred revenue, Breithaupt made it clear: “We are not guided by deferred revenue”, the guidelines are based on two quarters of visibility and communication with major customers; Ullal added that 2027 will have a large number of new products and new use cases.
Q4 and Annual Guidelines
Q4 guidance: Revenue of $3.9 billion to $4 billion (Bloomberg's agreed estimate of $3.57 billion), non-GAAP gross profit margin of 63% — 64%, non-GAAP operating margin 48% — 49% (agreed estimate of 48.5%), non-GAAP diluted EPS of $1.26 to 1.30 (agreed estimate of $1.139), diluted shares of about 1.28 billion shares, and a tax rate of about 21%. According to this estimate, revenue for the full year of 2026 was approximately US$13.3 billion (+48%), the fourth increase in the year (previously US$12.6 billion, with a consensus forecast of 12.7 billion dollars).
Regarding the big customer structure (Melius's Ben Reitzes asked the two 10% + customers mentioned earlier), Ullal stated that “diversification is the key word”: Microsoft and Meta are still “the closest 10% customers”, and it is expected that at least 1 new, possibly 2 10% customers will be added in 2026. “I feel like there are at least 3, maybe 4”, which ultimately depends on shipping capacity. Regarding the front-end network cycle (TD Cowen question), she answered “sine wave” alternately on the front-end/back-end: Currently, the back-end AI network is generally back-end, and the front-end will follow; 10% of customers concentrated on completing hundreds of G front-end refreshes this year, which may reverse next year.
Full Meeting Minutes
Company executives attending the conference: Jayshree Ullal (Chairman and CEO), Chantelle Breithaupt (Senior Vice President and Chief Financial Officer), Ken Duda (Founder, President and CTO), Todd Nightingale (President and COO), Rudolph Araujo (Head of Investor Relations, moderator)
Questioning analysts (in order of questions): Michael Ng (Goldman Sachs Research), Meta Marshall (Morgan Stanley Research), Ben Reitzes (Melius Research), George Notter (Wolfe Research), Amit Daryanani (Evercore ISI), Tim Long (Barclays Research), David Vogt (UBS Research), Simon Leopold (Raymond James), Joseph Cardoso (J.P. Morgan Chase Research), James Fish (Piper Sandler), Tal Liani (Bank of America Research), Ryan Koontz (Needham & Company), Karl Ackerman (BNP Paribas Research), Jin Marcote (Deutsche Bank Research), Daniel Zhu ( Bernstein), Sean O'Loughlin (TD Cowen), Antoine Chkaiban (New Street Research), Ben Bollin (Cleveland Research), Mehdi Hosseini (Susquehanna)
Presenter and opening
The moderator opened the meeting and prompted forward-looking statements and safe harbor statements. Rudolph Araujo introduced the participating management, explaining that the discussion was based on non-GAAP (excluding equity incentives, amortization of intangible assets, etc.), the FY2027 non-GAAP guidelines and GAAP caliber reconciliation tables “cannot be provided without unreasonable effort”, and then handed over to Jayshree Ullal.
Jayshree Ullal (Chairman and CEO) Preliminary Statement (Part 1)
Ullal announced revenue of US$3.66 billion for the quarter, an increase of 58.5% year over year, “this is the biggest year-on-year revenue increase since we became a listed company.” She focused on the results of the Arista 2.0 strategy over the past three years: moving from data centers to “data centers of data” (centers of data), covering clients to branches, campuses, data centers to clouds and AI centers with a single EOS operating system (5A principles: agility, automation, usability, analytics, AI); providing agentic (agentic) companion functions for automation and conversational troubleshooting based on a network data lake architecture. The “network-centric” acceleration in 2026 has become a trend: customers are increasingly “connecting” the back end and front end of AI data centers. The company was rated as a leader in innovation in the Forrester data center portfolio and is in a leading position in the Gartner campus's two magic quadrants for wired, wireless and data centers.
Regarding structural changes in the AI industry, she said that the industry is shifting from a “computation-centered vertical stack” to an “overall network-centered” approach: the diversity of heterogeneous models and accelerators requires an open network platform; the EtherLink portfolio has at least 25 switches for scale-up, scale-out, and scale-across, optimized to improve XPU utilization, task completion time, and token latency. At the OCP conference in October, the company showcased a rack-level scale-up design developed in collaboration with AMD, Arm, Broadcom, D-Matrix, Meta (MTIA), Microsoft (Maia), Qualcomm, etc., including orthogonal backplanes, high-density cable backplanes, and heterogeneous cross-rack designs; scale-up testing began in 2027, and the Ethernet port density in the rack increased to 144, about 1,000 connections across the rack, and mass production in 2028. The industry is moving from centralized training computation to distributed reasoning, prefill, and decode, putting higher demands on network intelligence; well-designed networks are becoming the “cornerstone of optimal token efficiency”.
AI customer side: “When we first started this AI journey, we were talking about 4-5 AI customers, about 50,000 to 100,000 NVIDIA GPUs, and in fact most of the networks were tied to proprietary systems such as InfiniBand. In 2026, we are in the AI supercycle, and there are already more than 100 Ethernet AI customers. By the end of 2027, we anticipate more than 10 gigawatts of deployment, spanning more than 10 different types of compute accelerators. What a significant inflection point within three years.” She also introduced Arista AVA (Autonomous Virtual Assistant) through MCP (Model Context Protocol) to open a programmable interface, which supports AI agents to interact with the network stack and correlate events across domains.
Ken Duda (Founder, President and CTO) speaks
Duda talks about the two-sided impact of AI on cybersecurity. The good side: AVA is driven by real-time data and can continuously monitor, troubleshoot, and automatically isolate the root cause between the campus, WAN, data center, and cloud. The bad side: AI is being used to find and exploit security breaches, and “true cybersecurity has become a necessity, not an option”. Over the past few months, the company has collaborated with Anthropic, Google, OpenAI, etc. to integrate the AI security capabilities of cutting-edge models into existing software security pipelines, and was invited to participate in AI-driven vulnerability discovery and assessment cooperation projects as a key infrastructure provider (original article “Project Class Win”, suspected to be a “Glasswing” project, see the checklist).
EOS's security architecture is based on four principles: first, a single high-quality operating system — the 20-year CVE count remains among the lowest in the industry (less than 100); second, single-component defects do not affect forwarding and state sharing architectures to resist horizontal movement through process isolation; third, seamless upgrades — affected components can be fixed on target without a maintenance window; fourth, the CloudVision compliance dashboard tracks security notices in real time. In addition, the company provides zero trust network (ZTN) capabilities: MSS (macro-segmentation service) to achieve dynamic security control between users, devices, and applications; NDR (network detection and response) services actively detect complex threats; and protect data center traffic with wire-speed, full-port native link encryption, with zero performance loss.
Todd Nightingale (President and COO) speaks
Nightingale introduced the rise in corporate business and three cases: First, an Asian network service company (the speed, scale, and resilience of the strategy shifts to the AI era) chose Arista to replace an existing supplier with long-term quality issues and deploy a 100G/400G platform based on the EVPN-VXLAN architecture, which will be the front-end network for its future AI deployment; second, in the medical field, an important hospital group selected Arista after a detailed zero-trust evaluation. The goal is to protect sensitive patient data with advanced network segmentation and seamlessly connect to the cloud. Arista uses R3 series backbone switches to centralize routing, deploy CloudVision as a Service and automation framework, and integrate MSS to achieve dynamic automation of the entire system; third, the hotel industry, an award-winning regional resort and casino unify their data center and campus architecture with Arista route switches and park access switches to support mission-critical applications such as casino game hall networks. What the three cases have in common is the quality and reliability of EOS, the seamless integration with the customer security ecosystem, and the TAC support team “escorting” the enterprise's migration.
Preliminary remarks by Jayshree Ullal (Part 2: Guidelines)
Ullal thanked the supply chain and manufacturing team for delivering under “industry-wide supply chain constraints and shortages”. The announcement: “We are now raising our 2026 guidance for the fourth time in the year to US$13.3 billion, representing 48% annual growth.” The company faces a “total serviceable market (TAM) of over $200 billion by 2030”, which is a multi-year growth opportunity. The 2027 guidance framework was presented for the first time: “Growth of 38% to 40%, or about US$18.5 billion. Among them, the AI network business is expected to be the fastest growing sector, doubling to about 7 billion US dollars; the cognitive park business achieved a healthy growth of 60%, far faster than the market, of about 2 billion US dollars in 2027.” She said that the company has surpassed the initial target of $10 billion this year, “both 2026 and 2027 will enter the double-digit (billion dollar) revenue range,” and achieve it using the Arista method of “innovation, quality, profitability, efficiency and zero debt.”
Chantelle Breithaupt (CFO) press release
Breithaupt said it achieved “three record highs of revenue growth, net profit, and EPS” (trifecta) this season, with an operating margin higher than 49%, while increasing investment in R&D, inventory and supply commitments ahead of demand. The company was included in the S&P 100 Index in September.
Financial details: Core revenue of approximately US$3.66 billion (+58.5%, guidance of US$3.3 billion); international revenue of US$682.5 million, accounting for 18.6% (23% in the previous quarter). The month-on-month change mainly stemmed from the geographical structure of major customer sales, with double-digit year-on-year growth in the three regions; gross profit margin of 64.0% (65.2% last year, 63.4% last quarter), supply chain efficiency partially offsetting the increase in product costs. The month-on-month improvement was driven by the terminal customer structure; operating expenses of US$545.3 million (accounting for 14.9% of revenue, 411 million in the previous quarter); R&D accounting for US$408.5 million (accounting for US$408.5 million in the previous quarter) 11.2%, $278.1 million in the previous quarter), plus liquid cooling, high power density switching and AI software; sales and marketing of US$110.9 million (3.0%, compared with 3.6% in the previous quarter); G&A of US$25.9 million (0.7%). Operating profit of US$1.8 billion (accounting for 49.1%); other income and expenditure was positive at US$138.6 million, effective tax rate of 20.9%; net profit of US$1.5 billion (accounting for 41.9% of revenue); diluted shares of 1,278 million shares, diluted EPS of US$1.20 ($0.77, +55.8% in the same period last year).
Balance sheet: Approximately US$14.4 billion in cash and marketable securities (US$13.3 billion at the end of the previous quarter); US$817.9 million remaining in the US$1.5 billion repurchase plan approved in May 2025; operating cash flow of approximately US$1.2 billion, benefiting from profit performance and intentional preparation to meet demand; DSO 59 days (68 days in the previous quarter); inventory turnover ratio 1.6 (1.7 billion in the previous quarter); inventory level fluctuations are expected to continue, and there may be a seasonally high inventory quarter affecting operating cash flow; US$11.4 billion in the previous quarter ($9.7 billion in the previous quarter) Mainly chip procurement required for new business and AI deployment; deferred revenue of about US$6.8 billion (6.9 billion in the previous quarter), mainly due to product delays, which reduced product delays by 86 million US dollars month-on-month — new product climbing, new customers and new use cases (including AI) brought more customer-specific acceptance terms and deferred fluctuations. “The deferred balance may fluctuate significantly quarterly independently of the underlying business”; 48 days of accounts payable (57 days in the previous quarter); capital expenditure of US$100.7 million, and the Santa Clara expansion project was completed within a few months.
Guidelines: Q4 revenue of US$3.9 billion to US$4 billion (corresponding to a median value of approximately US$13.3 billion for the full year of 2026); Q4 gross profit margin of 63% — 64%, operating margin of 48% — 49%; diluted EPS of 1.26 to 1.30 US dollars (approximately 1.28 billion diluted shares), with a tax rate of about 21%. 2027: Revenue growth of 38% to 40% (median value of about US$18.5 billion), gross profit margin 62% to 64%, operating profit margin of about 49%, and tax rate of about 21%.
Q & A session
Michael Ng (Goldman Sachs Research): Is “>10 gigawatts, 10 or more accelerators” an industry number or an Arista-specific caliber? How can computational diversification expand Arista's networking opportunities?
Ullal: AI is shifting from computing-centered to network-centric; the company supports three types of AI networks: scale-up, scale-out, and scale-across.” Our guess is that by the end of 2027 — 2027 is an exponential climbing year — more than 10 gigawatts will be deployed.” Some gigawatts have been added, but in the scale-across mode, both gigawatts can also be included. The company has already reached 4-5 types of XPUs at Scale-Across, and will add about 5 more types in 2027 as the scaling up progresses.
Meta Marshall (Morgan Stanley Research Department): What are the compositional changes in the three types of scale-across/scale-out/scale-up businesses, and the 2027 proportional outlook?
Ullal: The 2027 scale-up is still in early testing, and “does not constitute a substantial figure”; therefore, 2027 is mainly scale-out and scale-across: scale-across “at least one-third, probably higher in 2027”. This year, the two categories are roughly “seven or three open”, and may drop 70% and rise 30% next year. Arista is highly differentiated in Scale-Across with the 7800 platform — “When you connect to millions of XPUs, there is no room for fault tolerance for safety, performance, or reliability”. Currently, most of the demand is for back-end AI networks; as inference grows, the frontend and back-end boundaries become thinner, and the front-end will be re-examined.
Ben Reitzes (Melius Research): Ask if the two 10% + big customers mentioned earlier are in place and will they continue in 2027?
Ullal: “Diversification is the key word” - multiple accelerators, multiple models, and a diverse 10% customer.” Our closest partner is M&M (Microsoft and Meta), and we fully expect our top 10% of customers to continue with us in 2026, and hope for longer. We expect to add at least 1 or maybe 2 more. It depends a lot on Todd and the supply chain team being able to ship enough products—we left them until December 31st.” “We feel really good about at least 3, maybe 4 (10% customers),” she added.
George Notter (Wolfe Research): Does the sharp jump in R&D include one-off projects and is this the new normal?
Duda: Investing in new opportunities for AI-oriented network systems — Success with hyperscale AI companies depends on partner capabilities, and partner capabilities are based on understanding each company's unique needs and constraints; the introduction of liquid cooling as a new technology requires companies to carry out major infrastructure and technology upgrades.
Breithaupt: Maintain the framework of R&D accounting for 9% to 11% of revenue, which will peak in individual quarters due to specific investments; more detailed guidance for 2027 will be provided later.
Ullal (intermission): Prototype costs were really high this quarter—Duda: It's because of the opportunities in the hyperscale AI field.
Amit Daryanani (Evercore ISI): After the industry switched to a network-centric architecture, did Arista get involved in system design earlier and did network content density change? Also, with 38% to 40% confidence in the implementation of the guidelines in the supply chain, how much is already in the bag and how much still needs to be implemented?
Ullal: In the past, it was mainly involved in scale-out and scale-across; as cooperation with new accelerators deepened, the company understood its behavior at an earlier stage in the cycle — close cooperation with Cerebras and work with Microsoft (Maia) for several years have enabled the company to tune buffering, software performance, and maximize efficiency; these scale-up design results have been reflected in scale-out and scale-across.
Breithaupt: “I'm confident in the guidance, and I'm confident in the supply chain.” Internally is the supply chain team's excellent execution every day; outside is extensive communication and agreements with key suppliers — which is why procurement commitments are rising. In terms of visibility, “as usual, it's two quarters, and the latter two quarters need to be settled”, but with the flexibility of the product portfolio, the company knows where to bet.
Tim Long (Barclays Research): Q: The park and branches are growing at a high rate this year and will guide 60% next year — why so much faster than the industry? Has it been a trend for years?
Duda: “The corporate market is coming to us.” Over the past few years, almost all major enterprises have viewed campuses and branches as mission-critical networks; security incidents have accelerated this transformation, and customers need not only maximum reliability, but also the ability to quickly patch and adjust infrastructure — EOS's highly differentiated ability to have almost zero downtime is at the core of share growth.
Nightingale: The park's equipment renewal cycle is very long. It will take time for customers to understand Arista's advantages and adopt it (the pace of depreciation and upgrade). “This is expected to be a multi-year cycle”.
David Vogt (UBS Research): How do you understand these two metrics in the context of the 38% to 40% guideline when procurement promises increased by about 2 billion dollars month-on-month and deferred revenue declined?
Breithaupt: All three are designed to serve customer needs. A multi-year memory agreement has been added to the procurement commitment (to be seen in 10-Q), and is expected to rise slightly in Q4 to prepare for the 2027 guidance. Deferred revenue “had the largest flow in and out, and the balance remained flat”, while achieving 58.5% revenue growth. “This is a sign that demand is very healthy”; the year-end deferred balance is expected to be higher than at the end of last year.
Simon Leopold (Raymond James): Ask about the 2026 baseline for the 2027 campus and AI goals; and what does CPO (co-package optics) /near-package optics mean for Arista in the evolution of switches?
Breithaupt: The baseline is the 2026 guideline of US$13.3 billion, of which AI is about US$3.5 billion and the park is about US$1.25 billion.
Ullal: Optics is at the core of Arista's strategy (led by chief architect Andy Bechtolsheim, original text transcribed as “Andy Bessen”, see checklist): “Copper cables are used when they can be used on copper cables, and pluggable optics and lasers beyond a radius of 1-2 meters are the most widely deployed and accepted solutions.” Andy is an OSFP pioneer; new coherent optics/LPO solutions can achieve long ranges without DSP; additional miniaturized pluggable optics supported by more than 150 vendors are being promoted for higher panel densities.” We think the world will continue to be dominated by pluggable optics for the foreseeable future.” At the same time, in terms of higher speed, near-package optics/open CPO will be added — Arista is promoting open CPO, using pluggable optical engines and picobyte (original text vague) to achieve module pre-testing. “It is expected to enter the market between 2028 and 2030, which is not imminent. But at the same time, long live pluggable optics.”
Joseph Cardoso (J.P. Morgan Chase Research Department): What are the underlying assumptions that will guide the growth rate of AI and non-campus businesses in China in 2027 — high base, conservative supply, or others?
Ullal: “When some businesses grow by 60% or even 100%, and you say overall growth of 38% to 40%, obviously some businesses aren't that fast, right? We are very confident that all businesses will grow in double digits, some more, some less.” All sectors — AI, enterprises, professional cloud service providers, front-end networks — performed well.” At the beginning of 2026, we are still worried about whether the front end will grow” — the front end and back end are in a cyclical seesaw.” I'm not going to interpret these numbers as conservative. They are far above what we had anticipated. Please enjoy it and hope they'll be even better.”
James Fish (Piper Sandler): The team has always only given 1-2 quarters of visibility and style prudence. Why did they give such strong guidance for 2027 so early? And where did the confidence in the 2028 scale-up come from?
Breithaupt: This strong guidance, given one year in advance, has been fully proven — Ken, Jayshree, and the team “first made themselves feel at ease”: two quarters of visibility, two years of customer relationships and technical conversations, “usually the second half of the year settled down like this, with no changes”.
Ullal: Scale-up has always been a tough market for us — it doesn't have all the scenarios where our software excels, but Arista has the strongest hardware ability to connect to any non-NVIDIA accelerator.” Just as we were staring at InfiniBand and saying 'of course it would be Ethernet', now we're eyeing proprietary UA Link or NVLink and saying 'of course it will be ESUN'.” This type of migration takes years; “although we have only two quarters of visibility, we have two years of work behind us”, so I'm confident in both the 2027 guidelines and 2028.
Tal Liani (Bank of America Research): While the guidelines were drastically raised, total deferral and current deferrals are slowing down month-on-month — is growth driven by new orders or deferred releases? Is deferred revenue losing its ability to predict?
Ullal: “You've broken down the requirements into a physics problem, and I don't have an answer. I'm not predicting or predicting delays. “Deferral is a strong function of new use cases and new products” — millions of dollars flow out after use cases are satisfied, and new products bring in millions of dollars; look at delays as an upward trend, “but delays will never grow indefinitely”. What matters is the big picture: “We have AI, enterprise, campus, and data center needs, and this was the best time in my 18-year career for the company.”
Breithaupt: Fluctuations in deferral are not quarterly entangled; very confident in demand and guidance; deferrals only reflect balance sheet accounts that occurred during the quarter.” “Reiterating what we said last quarter: the deferred balance at the end of this year is expected to be higher than at the end of last year — this shows that P&L did what it was supposed to do without relying on deferrals.”
Ryan Koontz (Needham & Company): How did the delivery model change after scale-up products introduced more optics and liquid cooling — whether the entire rack was sold by itself or by a partner? How will the product plan unfold?
Ullal: “We're still exploring ourselves.” The core of Arista's value at the rack level is still the switch, but the entire rack must be designed and then worked with different partners in different ways — some require the entire rack, some are delivered through manufacturing partners, and Arista only supplies the switch.” “Every book is hard to miss” — The company has three designs: orthogonal, cable backplane, and cross-rack, all of which are different, and will be disclosed separately in the next few quarters. One thing is clear: “A large number of switches are being shipped, and many are liquid cooled.”
Karl Ackerman (BNP Paribas Research): Does the 2027 guidance presuppose approval/financing of new electricity infrastructure? Is aggressive deferred revenue consumption required to achieve it?
Breithaupt: “We are not guided by deferred revenue.” The guidance is based on two quarters of visibility, conversations with the largest customers, and the operating rate of cloud and enterprise businesses.” But we're seeing very strong signs of demand.”
Ullal: Clearly added: “While we are not directing delays, 2027 will be as full of new products and new use cases as 2026.”
Jin Marcote (Deutsche Bank Research Division): After the SU-144 is mass-produced in 2028, how will the gross margin change compared to the existing switch mix? How many customer promises have been made?
Ullal: “Since listing, our long-term gross margin has been 60% — 65%. Now we're narrowing the range to 62% — 64%. We don't see any change.”
Daniel Zhu (Bernstein): Ask about Scale-Across's customer concentration - mainly cloud giants do simultaneous training, or do telecom operators and smaller customers and inference use cases?
Ullal: There are both: colo and smaller use cases will use smaller platforms such as 7280; larger customers prefer very large AI spines across data centers, requiring 576 port densities, terabytes, and multi-gigawatt deployments.” The revenue concentration is biased towards 7800, and the geographical concentration of multi-site deployments is biased towards 7280.”
Sean O'Loughlin (TD Cowen): Ask about the cyclical position of front-end networks — the market is paying attention to 800G and above back-ends. What stage is traditional front-end computing currently at?
Ullal: “It's like a sine wave.” When customers focus on the back-end, front-end investment decreases, and vice versa; when extended to six quarters to two years, the two will offset each other, but in a specific year they will focus on one side — 10% of customers concentrated on completing hundreds of G front-end refreshes in one year, which may reverse the next year.” Overall, we're seeing more back-end and AI networks now, and the front-end will keep up.”
Antoine Chkaiban (New Street Research): Ask about the commercialization of AVA — how to quantify this opportunity given the recent high profile infrastructure security incidents? Is it classified as product revenue or service revenue?
Duda: “Everything is still in its early stages.” AVA has just opened up customer trials as an alpha feature in the CloudVision service. The pricing model is being developed, and “there are currently no predictions or data to share”.
Ullal: But the prospects are promising — the EOS architecture is solid enough (“we can fix our bugs from 30 to 67, so we can superimpose AVA; others are playing a game with thousands of bugs that can't be traced”), and “we feel we can monetize AVA in the next few years.”
Ben Bollin (Cleveland Research): What is the current status of the delivery cycle (lead times) of enterprise-level and hyperscale businesses; and the progress and trend of companies seizing market share?
Nightingale: The enterprise side delivery period is short - the switch type used is not in the most advanced technology, the unit price is lower, and more buffer inventory can be kept; the delivery time for components and finished products in AI and hyperscale data centers is longer. In terms of marketing, the momentum was strong this year, and the webinars and events were well received. The largest customer conference in the company's history will be held in December; sales execution is “amazingly efficient” and will continue to expand next year.
Ullal: Added to Nightingale and sales leaders (original text vague), increased investment in the US and key industries and international markets over the past year, and the go-to-market investment over the past three years “is being realized”.
Mehdi Hosseini (Susquehanna): Is the supply of chips and storage all locked in the $18.5 billion 2027 guideline? There are opinions that storage supply will be tighter in 2027. Are there any downside risks in the guidelines?
Nightingale: The company is proud of its strategic partnership, has invested in large-scale long-term agreements in PCB production capacity and memory (as described above) other than silicon wafers, and “has a very high level of confidence in its ability to deliver supplies as predicted. Of course, we won't be placing orders five quarters in advance for the whole of next year—not every order has been placed.”
Concluding remarks
Rudolph Araujo announced the end of the conference call, reminding that the company had posted additional presentation materials on the Investor Relations website. Thank you for participating.