Arista Networks (ANET.US) Q3 Earnings Call: CEO Says More Than 100 Ethernet AI Customers, Entering AI Super Cycle; First-Time Guidance of US$18.5 Billion Revenue for 2027

Arista Networks (ANET.US) Q3 Earnings Call: CEO Says More Than 100 Ethernet AI Customers, Entering AI Super Cycle; First-Time Guidance of US$18.5 Billion Revenue for 2027 Beijing time early morning November 4 (after market close on November 3 US Eastern Time), cloud networking equipment vendor Arista Networks (ANET.US) released its third-quarter 2026 results for the period ended September 30, and held an earnings call after market close the same day. Third-quarter revenue was US$3.66 billion, up 20.5% quarter-over-quarter and up 58.5% year-over-year, the company's highest single-quarter year-over-year growth rate since its IPO, exceeding the company's guidance of approximately US$3.3 billion and also above the Bloomberg consensus estimate of US$3.33 billion. Non-GAAP gross margin was 64.0%, down 1.2 percentage points year-over-year and up 0.6 percentage points quarter-over-quarter, above the company's prior guidance of approximately 63%; non-GAAP operating margin was 49.1% (48.6% in the year-ago period); non-GAAP net income was US$1.53 billion, up 56% year-over-year; non-GAAP diluted EPS was US$1.20, up 55.8% year-over-year, above the Bloomberg consensus estimate of US$1.075. GAAP diluted EPS was US$1.22, up 82.1% year-over-year, higher than the non-GAAP figureaccording to GMTEight's breakdown based on the financial report, mainly because GAAP results included approximately US$134.5 million in strategic investment gains (excluded from non-GAAP). In addition, the company was added to the S&P 100 Index in September. On the share price front, the stock surged in after-hours trading following the results, rising as much as 12% at one point; as of publication, it was up nearly 7% in overnight trading. Market Focus and Management Response The focus this quarter was not on the Q3 results themselves but on the 2027 guidance management provided for the first time: revenue of approximately US$18.5 billion (growth of 38%40%), significantly above the Bloomberg consensus estimate of US$16.4 billion; of which AI networking (AI fabrics) is targeted at approximately US$7 billion, doubling from the 2026 baseline of US$3.5 billion, and cognitive campus is targeted at approximately US$2 billion, up 60% (2026 baseline of US$1.25 billion); gross margin guidance of 62%64%, operating margin of approximately 49%. At the same time, the company raised its full-year 2026 revenue guidance for the fourth time this year to US$13.3 billion (+48% year-over-year). CEO Jayshree Ullal characterized this round of demand as an AI super cycle: "In 2026, we are in the AI super cycle, with more than 100 Ethernet AI customers. By the end of 2027, we forecast more than 10 gigawatts of deployments across more than 10 different types of compute accelerators. What a remarkable inflection point in just three years." She recalled that when the company started three years ago, it had only 45 AI customers, approximately 50,000100,000 NVIDIA GPUs, and networking was mostly tied to proprietary systems such as InfiniBand. Regarding concerns about supply assurance (Mehdi Hosseini of Susquehanna pressed on whether chips and memory had all been locked in), President and COO Todd Nightingale responded that the company has "made large-scale long-term strategic agreement investments in silicon, PCB capacity and memory," and has "very high confidence" in delivering supply in line with the 2027 guidance, but "we don't place orders five quarters ahead for all orders for the entire next yearnot every order is in place yet." CFO Chantelle Breithaupt added that the company has demand visibility for the next two quarters, and the increase in purchase commitments to US$11.4 billion is precisely a reflection of confidence. First-Time 2027 Guidance: AI Doubling to US$7 Billion, Campus Up 60% Management disclosed its 2027 framework for the first time: revenue of approximately US$18.5 billion (+38%40%), with the AI networking business doubling to approximately US$7 billion as the fastest-growing segment; campus business of approximately US$2 billion (+60%), "far faster than the market"; gross margin of 62%64%, operating margin of approximately 49%, tax rate of approximately 21%. The corresponding 2026 baseline: full-year revenue of US$13.3 billion, of which AI is approximately US$3.5 billion and campus approximately US$1.25 billion. Ullal explained why 2027 total growth is lower than the AI segment's growth: "When some businesses are growing 60% or even 100%, and overall growth is 38%40%, obviously some businesses are not growing as fast. We are confident all businesses will maintain double-digit growth." She stressed this is not conservative: "Our starting guidance is already far above our past forecasts. Please enjoy it, and hopefully it will get even better." The company also raised its estimate of the 2030 total addressable market (TAM) to "more than US$200 billion." AI Networking: More Than 100 Customers, 10 Gigawatts of Deployments and the Scale-up Timeline Ullal said the company supports three types of AI networking: scale-up (within the rack), scale-out (within the cluster) and scale-across (across data centers). Scale-out and scale-across remain the main revenue contributors; this year the split between the two was roughly "70/30," and in 2027 scale-across will account for "at least one-third, possibly higher," with the flagship product being the 7800 AI Spine. On scale-up, the company in October at the Open Compute Project (OCP) conference showcased a rack-level scale-up design, working with partners including AMD, Arm, Broadcom Inc., d-Matrix, Meta, Microsoft Corporation and Qualcomm, and based on the Ethernet Scale-Up Networking (ESUN) standard introduced the Etherlink SU-144 architecture supporting 144 accelerators per rack and scaling across racks to 1,024 XPUs; scale-up trials will begin in 2027, with volume production in 2028, at which point in-rack Ethernet speeds will rise to 144 ports and cross-rack connections to approximately 1,000 XPUs. She expects scale-up revenue contribution in 2027 to "not constitute a meaningful number," with contribution beginning in 2028. Regarding the basis for the 10 gigawatts figure (Michael Ng of Goldman Sachs Group, Inc. asked whether it was an industry number or a company number), Ullal responded: "Our guess is that by the end of 2027with the exponential ramp occurring mainly in 2027more than 10 gigawatts will be deployed," including both new gigawatt-scale deployments and, under the scale-across model, gigawatt-scale capacity utilizing existing data centers; the company currently reaches 45 types of XPUs in scale-across, and with the advancement of scale-up, approximately 5 more will be added in 2027. Enterprise and Campus: Security-Driven, Multi-Year Cycle President and COO Todd Nightingale introduced three representative wins this quarter: an Asian network services company that chose Arista due to a strategic shift toward speed, scale and resilience in the AI era (a 100G/400G platform based on EVPN-VXLAN, which will serve as the front-end network for its AI deployments); a leading hospital group that selected Arista after a rigorous zero-trust assessment, using MSS (Macro Segmentation Service) to achieve dynamic security segmentation; and a regional resort and casino that updated its mission-critical network with Arista routing and switching plus campus access switches. Founder and CTO Ken Duda placed security at the center of the AI narrative: the company works with Anthropic, Google, OpenAI and others to embed frontier models' AI security capabilities into the vulnerability discovery process (such as "Glasswing"-type projects; the original transcript is unclear, see verification table); the EOS single-operating-system architecture keeps its CVE count "under 100, among the lowest in the industry for two decades," and Ullal said it is "10 to 50 times better than alternatives." Duda attributed the campus high growth to "the enterprise market coming to us"almost every major enterprise regards campus and branch as mission-critical networks; Nightingale added that campus equipment refresh cycles are long, and customer awareness and switching take time, "we expect this to be a multi-year cycle." Financial Details: R&D Stepped Up Significantly, Inventory Build and Commitment Expansion CFO Chantelle Breithaupt disclosed: international revenue was US$682.5 million, or 18.6% (23% last quarter), with the quarter-over-quarter change mainly due to the geographic mix of large-customer sales, and all three major regions achieved double-digit year-over-year growth; operating expenses were US$545.3 million (14.9% of revenue), of which R&D expenses were US$408.5 million, or 11.2% of revenue (US$278.1 million last quarter), with the increase directed toward liquid cooling, high-power-density switching and AI softwarein response to Wolfe Research's question about the R&D jump, Duda explained it as infrastructure upgrades driven by collaboration requirements with hyperscale AI customers and the introduction of liquid cooling technology, and Breithaupt maintained the framework of R&D at 9%11% of revenue; sales and marketing expenses were 3.0%. Operating cash flow was approximately US$1.2 billion; inventory was US$3.2 billion (US$2.5 billion last quarter); purchase commitments were US$11.4 billion (US$9.7 billion last quarter), including newly signed multi-year memory agreements; deferred revenue was approximately US$6.8 billion (US$6.9 billion last quarter), with product deferred revenue decreasing by US$86 million quarter-over-quarter. In response to UBS's question about the relationship between deferred revenue and the 38%40% guidance, Breithaupt responded: "Deferred revenue had its largest in-and-out flow this quarter while the balance stayed flat, and at the same time we achieved 58.5% revenue growththis is a sign of very healthy demand"; she also expects the year-end deferred revenue balance to be higher than last year-end. To Bank of America's Tal Liani's pointed question about "orders and deferred revenue being decoupled," Ullal said bluntly: "You want to break demand down into a physics problem, and I don't have an answer... What matters is the big picture: we have demand in AI, enterprise, campus and data center, and this is the best position the company has been in during my 18-year career." In response to BNP Paribas's question about whether the 2027 guidance depends on newly built power facilities or drawing down deferred revenue, Breithaupt made clear: "We do not guide based on deferred revenue"; the guidance is based on two quarters of visibility and communications with large customers; Ullal added that 2027 will bring a large number of new products and new use cases. Q4 and Full-Year Guidance Q4 guidance: revenue of US$3.94.0 billion (Bloomberg consensus US$3.57 billion), non-GAAP gross margin of 63%64%, non-GAAP operating margin of 48%49% (consensus 48.5%), non-GAAP diluted EPS of US$1.261.30 (consensus US$1.139), diluted shares of approximately 1.28 billion, tax rate of approximately 21%. Based on this, full-year 2026 revenue is approximately US$13.3 billion (+48%), the fourth raise this year (previously US$12.6 billion, consensus US$12.7 billion). On large-customer structure (Ben Reitzes of Melius asked about the two 10%+ customers previously mentioned), Ullal said "diversification is the key word": Microsoft Corporation and Meta remain the "closest 10% customers," and the company expects to add at least 1, possibly 2, 10%-level customers in 2026, "feeling like at least 3, maybe 4," ultimately depending on shipment capability. On the front-end networking cycle (TD Cowen's question), she answered by describing front-end/back-end as alternating "sine waves": currently the overall tilt is toward back-end AI networking, with front-end to follow later; this year one 10% customer concentrated a several-hundred-gig front-end refresh, and next year it may reverse. Full Conference Transcript Company management participating: Jayshree Ullal (Chair and CEO), Chantelle Breithaupt (SVP and CFO), Ken Duda (Founder, President and CTO), Todd Nightingale (President and COO), Rudolph Araujo (Head of Investor Relations, moderator) Analysts asking questions (in order): Michael Ng (Goldman Sachs Group, Inc. 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 Group AG Research), Simon Leopold (Raymond James), Joseph Cardoso (JPMorgan Research), James Fish (Piper Sandler), Tal Liani (Bank of America Corp Research), Ryan Koontz (Needham & Company), Karl Ackerman (BNP Paribas Research), Jin Marcote (Deutsche Bank Aktiengesellschaft Research), Daniel Zhu (Bernstein), Sean O'Loughlin (TD Cowen), Antoine Chkaiban (New Street Research), Ben Bollin (Cleveland Research), Mehdi Hosseini (Susquehanna) Moderator and Opening The moderator opened the meeting and noted forward-looking statements and the safe harbor statement. Rudolph Araujo introduced the participating management, explaining that the discussion is based on a non-GAAP basis (excluding stock-based compensation, amortization of intangible assets, etc.), and that a reconciliation of the FY2027 non-GAAP guidance to GAAP "cannot be provided without unreasonable effort," then handed over to Jayshree Ullal. Jayshree Ullal (Chair and CEO) Prepared Remarks (Part One) Ullal announced quarterly revenue of US$3.66 billion, up 58.5% year-over-year, "the largest year-over-year revenue growth since we became a public company." She highlighted the results of the Arista 2.0 strategy over three years: moving from data centers to "centers of data," covering client to branch, campus, data center to cloud and AI centers with a single EOS operating system (the 5A principles: Agile, Automated, Available, Analytics, AI); and providing agentic companion capabilities for automation and conversational troubleshooting based on a network data lake architecture. In 2026, the "networking-centric" acceleration has become a trend: customers are increasingly "chaining together" the back end and front end of AI data centers. The company received a Forrester leader rating for data center portfolio innovation and holds leadership positions in both the Gartner Magic Quadrant for Campus Wired and Wireless and the Magic Quadrant for Data Center. On the structural changes in the AI industry, she said the industry is shifting from "compute-centric vertical stacks" to a "holistic networking-centric" approach: the diversity of heterogeneous models and accelerators requires open networking platforms; the Etherlink portfolio has at least 25 switches for scale-up, scale-out and scale-across, optimized to improve XPU utilization, job completion time and token latency. At the October OCP conference, the company showcased rack-level scale-up designs developed with AMD, Arm, Broadcom Inc., d-Matrix, Meta (MTIA), Microsoft Corporation (Maia), Qualcomm and others, including orthogonal backplanes, high-density cable backplanes and heterogeneous cross-rack designs; scale-up trials begin in 2027, with in-rack Ethernet port density rising to 144 and approximately 1,000 cross-rack connections, and volume production in 2028. The industry is moving from centralized training compute to distributed inference, prefill and decode, placing higher demands on network intelligence; well-designed networks are becoming "the cornerstone of optimal token efficiency." On the AI customer side: "When we first began this AI journey, we were talking about 45 AI customers, roughly 50,000 to 100,000 NVIDIA GPUs, and networking was mostly tied to proprietary systems such as InfiniBand. In 2026, we are in the AI super cycle, with more than 100 Ethernet AI customers. By the end of 2027, we forecast more than 10 gigawatts of deployments across more than 10 different types of compute accelerators. What a remarkable inflection point in three years." She also introduced Arista AVA (Autonomous Virtual Assistant), which opens programmable interfaces via MCP (Model Context Protocol), enabling AI agents to interact with the network stack and correlate events across domains. Ken Duda (Founder, President and CTO) Remarks Duda discussed AI's dual impact on network security. The good side: AVA is driven by real-time data and can continuously monitor and troubleshoot across campus, WAN, data center and cloud, and automatically isolate root causes. The bad side: AI is being used to discover and exploit security vulnerabilities, "real cybersecurity has become a necessity, not a nice-to-have." Over the past several months, the company has worked with Anthropic, Google, OpenAI and others to integrate frontier models' AI security capabilities into its existing software security pipeline, and has been invited as a critical infrastructure provider to participate in AI-driven vulnerability discovery and assessment collaboration projects (the original says "Project Class win," suspected to be a "Glasswing"-type project, see verification table). EOS's security architecture is based on four principles: first, a single high-quality operating systemCVE count has remained among the lowest in the industry for two decades (under 100); second, limiting the blast radius of vulnerabilities through process isolation, so a single component flaw does not affect forwarding, and a state-sharing architecture resists lateral movement; third, seamless upgradestargeted fixes to affected components without maintenance windows; fourth, the CloudVision compliance dashboard tracks security advisories in real time. In addition, the company offers Zero Trust Networking (ZTN) capabilities: MSS (Macro Segmentation Service) enables dynamic security controls among users, devices and applications; NDR (Network Detection and Response) services proactively detect sophisticated threats; and link encryption native to all ports at wire speed protects traffic leaving the data center with zero performance penalty. Todd Nightingale (President and COO) Remarks Nightingale introduced the momentum in the enterprise business and three cases: first, an Asian network services company (strategically shifting toward the speed, scale and resilience of the AI era) chose Arista to replace an incumbent supplier with long-standing quality issues, deploying a 100G/400G platform based on an EVPN-VXLAN architecture that will serve as the front-end network for its future AI deployments; second, in healthcare, a major hospital group selected Arista after a detailed zero-trust assessment, aiming to protect sensitive patient data with advanced network segmentation and seamlessly connect to the cloud; Arista centralized routing with R3 series spine switches, deployed CloudVision as a Service and an automation framework, and integrated MSS to achieve dynamic automation across the system; third, in hospitality, an award-winning regional resort and casino unified its data center and campus architecture with Arista routing and switching and campus access switches, supporting mission-critical applications such as the casino gaming floor network. What the three cases have in common is the quality and reliability of EOS, seamless integration with customers' security ecosystems, and the TAC support team "escorting" enterprise migrations. Jayshree Ullal Prepared Remarks (Part Two: Guidance) Ullal thanked the supply chain and manufacturing teams for delivering amid "industry-wide supply chain constraints and shortages." She announced: "We are now raising our annual 2026 guidance for the fourth time this year to US$13.3 billion, representing 48% annual growth." The company's "total addressable market (TAM) exceeds US$200 billion by 2030," a multi-year growth opportunity. She provided the first 2027 guidance framework: "Growth of 38% to 40%, approximately US$18.5 billion. Of which, the AI networking business is expected to be the fastest-growing segment, doubling to approximately US$7 billion; the cognitive campus business achieving healthy growth of 60%, far faster than the market, at approximately US$2 billion in 2027." She said the company has already surpassed its initial US$10 billion target this year, and "both 2026 and 2027 will be in the double-digit (tens of billions of dollars) revenue range," achieved the Arista way with "innovation, quality, profitability, efficiency and zero debt." Chantelle Breithaupt (CFO) Prepared Remarks Breithaupt said the quarter achieved a "trifecta of record revenue growth, net income and EPS," with operating margin above 49%, while stepping up investment in R&D, inventory and supply commitments ahead of demand. The company was added to the S&P 100 Index in September. Financial details: core revenue of approximately US$3.66 billion (+58.5%, guidance US$3.3 billion); international revenue of US$682.5 million, or 18.6% (23% last quarter), with the quarter-over-quarter change mainly due to the geographic mix of large-customer sales, and all three major regions growing double digits year-over-year; gross margin of 64.0% (65.2% a year ago, 63.4% last quarter), with supply chain efficiency partly offsetting higher product costs, and the quarter-over-quarter improvement driven by end-customer mix; operating expenses of US$545.3 million (14.9% of revenue, US$411 million last quarter); R&D of US$408.5 million (11.2%, US$278.1 million last quarter), with additional investment in liquid cooling, high-power-density switching and AI software; sales and marketing of US$110.9 million (3.0%, 3.6% last quarter); G&A of US$25.9 million (0.7%). Operating income of US$1.8 billion (49.1%); other income/expense was a positive US$138.6 million, effective tax rate 20.9%; net income of US$1.5 billion (41.9% of revenue); diluted shares of 1.278 billion, diluted EPS of US$1.20 (US$0.77 in the year-ago period, +55.8%). Balance sheet: cash and marketable securities of approximately US$14.4 billion (US$13.3 billion at the end of last quarter); US$817.9 million remaining available under the US$1.5 billion buyback program approved in May 2025; operating cash flow of approximately US$1.2 billion, benefiting from earnings performance and deliberate inventory builds to meet demand; DSO of 59 days (68 days last quarter); inventory turns of 1.6 (1.7 last quarter); inventory of US$3.2 billion (US$2.5 billion last quarter), with inventory level fluctuations expected to continue and possible seasonal high-inventory quarters affecting operating cash flow; purchase commitments of US$11.4 billion (US$9.7 billion last quarter), mainly chip purchases required for new business and AI deployments; deferred revenue of approximately US$6.8 billion (US$6.9 billion last quarter), mainly product deferred, with product deferred decreasing by US$86 million quarter-over-quarternew product ramps, new customers and new use cases (including AI) bring more customer-specific acceptance terms and deferred fluctuations, "the deferred balance may fluctuate significantly quarter to quarter independently of the underlying business drivers"; days payable outstanding of 48 days (57 days last quarter); capital expenditures of US$100.7 million, with the Santa Clara expansion completing within a few months. Guidance: Q4 revenue of US$3.94.0 billion (corresponding to full-year 2026 of approximately US$13.3 billion at or above the midpoint); Q4 gross margin of 63%64%, operating margin of 48%49%; diluted EPS of US$1.261.30 (diluted shares of approximately 1.28 billion), tax rate of approximately 21%. 2027: revenue growth of 38%40% (approximately US$18.5 billion at the midpoint), gross margin of 62%64%, operating margin of approximately 49%, tax rate of approximately 21%. Q&A Session Michael Ng (Goldman Sachs Group, Inc. Research): Asked whether ">10 gigawatts, more than 10 types of accelerators" is an industry number or an Arista-specific figure, and how compute diversification expands Arista's networking opportunity? Ullal: AI is shifting from compute-centric to networking-centric; the company supports three types of AI networking: scale-up, scale-out and scale-across. "Our guess is that by the end of 20272027 being the year of exponential rampdeployments will exceed 10 gigawatts." Part is new gigawatts, but under the scale-across model existing gigawatts can also be included. The company already reaches 45 types of XPUs in scale-across, and with the advancement of scale-up, approximately 5 more will be added in 2027. Meta Marshall (Morgan Stanley Research): Asked about the mix shift among scale-across/scale-out/scale-up and the 2027 ratio outlook? Ullal: In 2027, scale-up will still be in early trials and "will not constitute a materially meaningful number," with contribution likely only in 2028; therefore 2027 will mainly be scale-out and scale-across: scale-across "at least one-third, possibly higher in 2027"; this year the two were roughly "70/30," and next year the 70% may decline and the 30% rise. Arista is highly differentiated in scale-across with the 7800 platform"when you are connecting millions of XPUs, there is no margin for error in security, performance or reliability." Currently demand is mostly for back-end AI networking; as inference grows, the boundary between front end and back end will blur, and the front end will be revisited. Ben Reitzes (Melius Research): Asked whether the two 10%+ large customers previously mentioned are in place and whether they will continue in 2027? Ullal: "Diversification is the key word"diverse accelerators, diverse models, and also diverse 10% customers. "Our closest partners are M&M (Microsoft Corporation and Meta), and we fully expect the existing top 10% customers to continue with us in 2026 and hopefully longer; we expect to add at least 1, possibly 2. That depends heavily on whether Todd and the supply chain team can ship enough productwe're giving them until December 31." She added: "We feel very good about at least 3, maybe 4 (10% customers)." George Notter (Wolfe Research): Asked whether the large quarter-over-quarter jump in R&D includes one-time items and whether it is the new normal? Duda: The investment targets new opportunities in AI networking systemssuccess with hyperscale AI companies depends on partner capabilities, and partner capabilities are built on understanding each company's unique needs and constraints; liquid cooling, as a new technology introduction, requires the company to undertake significant infrastructure and technology upgrades. Breithaupt: Maintaining the framework of R&D at 9%11% of revenue, with individual quarters peaking due to specific investments; more detailed 2027 guidance will be provided later. Ullal (interjecting): Prototype costs were indeed high this quarterDuda: precisely because of the opportunity in hyperscale AI. Amit Daryanani (Evercore ISI): Asked whether, as the industry shifts to networking-centric architectures, Arista is engaging earlier in system design and whether network content density is changing; and in the confidence behind the 38%40% guidance for supply chain execution, how much is already in the bag and how much still needs to be secured? Ullal: In the past we mainly participated in scale-out and scale-across; as collaboration with new accelerators deepens, the company understands their behavior earlier in the cycleclose collaboration with Cerebras and work with Microsoft Corporation (Maia) have been underway for several years, allowing the company to tune buffers, software performance and maximize efficiency; these scale-up design achievements are already reflected in scale-out and scale-across. Breithaupt: "I am confident in the guidance and confident in the supply chain." Internally, it is the supply chain team's excellent execution every day; externally, it is extensive communication and agreements with key supplierswhich is precisely why purchase commitments are rising. On visibility, "as usual it is two quarters, and the latter two quarters need to be secured," but with the flexibility of the product portfolio, the company knows where to place its bets. Tim Long (Barclays Research): Asked about campus and branch high growth this year and 60% guidance for next yearwhy far faster than the industry? Is it a multi-year trend? Duda: "The enterprise market is coming to us." Over the past few years, almost every major enterprise has regarded campus and branch as mission-critical networks; security incidents have accelerated this shift, with customers wanting both the highest reliability and the ability to patch and adjust infrastructure quicklyEOS's near-zero-downtime capability is highly differentiated and is core to share gains. Nightingale: Campus equipment refresh cycles are long, and it takes time for customers to understand Arista's advantages and adopt them (depreciation and upgrade cadence), "we expect this to be a multi-year cycle." David Vogt (UBS Group AG Research): Asked about purchase commitments rising by approximately US$2 billion quarter-over-quarter while deferred revenue is declining, and how to understand these two metrics in the context of the 38%40% guidance? Breithaupt: All three are designed to serve customer demand. Purchase commitments include newly signed multi-year memory agreements (to be seen in the 10-Q), and are expected to rise slightly in Q4 in preparation for the 2027 guidance. Deferred revenue this quarter had "the largest in-and-out flow while the balance stayed flat," while achieving 58.5% revenue growth, "this is a sign of very healthy demand"; the year-end deferred balance is expected to be higher than last year-end. Simon Leopold (Raymond James): Asked about the 2026 baselines for the 2027 campus and AI targets; and what CPO (co-packaged optics)/near-packaged optics mean for Arista in switch evolution? Breithaupt: The baseline is the 2026 guidance of US$13.3 billion, of which AI is approximately US$3.5 billion and campus approximately US$1.25 billion. Ullal: Optics is core to Arista's strategy (led by Chief Architect Andy Bechtolsheim, transcribed in the original as "Andy Bessen," see verification table): "Use copper where copper can be used; beyond a 12 meter radius, pluggable optics and lasers are the most widely deployed and accepted solution." Andy is a pioneer of OSFP; new coherent optics/LPO solutions can achieve long reach without DSPs; for higher faceplate density, additional miniaturized pluggable optics supported by more than 150 vendors are advancing. "We believe the world will remain primarily pluggable optics for the foreseeable future." At the same time, at higher speeds, near-packaged optics/open CPO will be complementaryArista is advocating open CPO, using pluggable optical engines and picobyte (original unclear) to enable module pre-testing, "expected to enter the market between 2028 and 2030, not imminent. But in the meantime, long live pluggable optics." Joseph Cardoso (JPMorgan Research): Asked about the underlying assumptions for slower growth in non-AI, non-campus businesses in the 2027 guidancehigh base, supply conservatism or something else? Ullal: "When some businesses are growing 60% or even 100%, and you say overall growth is 38%40%, obviously some businesses are not growing as fast, right? We are very confident all businesses will grow double digits, some more, some less." All segmentsAI, enterprise, specialized cloud providers, front-end networkingare performing well. "At the start of 2026 we were still worried about whether the front end would grow"front end and back end are a cyclical seesaw. "I wouldn't interpret these numbers as conservative. They are far above what we once forecast. Please enjoy them, and hopefully they will get even better." James Fish (Piper Sandler): Asked why the team, which has consistently given only 12 quarters of visibility and is conservative in style, is giving such strong 2027 guidance so early? And where does confidence in 2028 scale-up come from? Breithaupt: This strong guidance given a year early is well substantiatedKen, Jayshree and the team "first made themselves comfortable": two quarters of visibility, two years of customer relationships and technology conversations, "this is how the second half typically shapes up, no change." Ullal: Scale-up has always been a difficult market for usit doesn't have the full scope of our software advantages, but Arista has the strongest hardware capability in connecting all non-NVIDIA accelerators. "Just as we looked at InfiniBand back then and said 'of course it will be Ethernet,' today we look at proprietary UA Link or NVLink and say 'of course it will be ESUN.'" Such transitions take years; "even though visibility is only two quarters, we have two years of work behind us," hence full confidence in the 2027 guidance and 2028. Tal Liani (Bank of America Corp Research): Asked whether, with guidance raised significantly while total and current deferred revenue decelerate quarter-over-quarter, the growth driver is really new orders or deferred revenue release? Is deferred revenue losing its predictive power? Ullal: "You have turned demand into a physics problem, and I don't have an answer. I don't forecast or pre-announce deferred revenue. Deferred revenue is a strong function of new use cases and new products"millions of dollars flow out when use cases are satisfied, and millions flow in with new products; over the longer term deferred revenue is trending upward, "but deferred revenue will never grow indefinitely." What matters is the big picture: "We have demand in AI, enterprise, campus and data center, and this is the best position the company has been in during my 18-year career." Breithaupt: We don't obsess over deferred revenue fluctuations quarter by quarter; we are very confident in demand and guidance; deferred revenue merely reflects a balance sheet line item that occurs in the quarter. "To reiterate what we said last quarter: we expect the year-end deferred balance to be higher than last year-endwhich shows the P&L did what it needed to do without relying on deferred revenue." Ryan Koontz (Needham & Company): Asked how the delivery model changes as scale-up products incorporate more optics and liquid coolingselling complete racks ourselves or partner delivery? How will the product plan unfold? Ullal: "We ourselves are still figuring it out." Arista's core value at the rack level remains the switch, but it must design the entire rack and then work with different partners in different wayssome want complete racks, some take delivery through manufacturing partners with Arista supplying only the switch. "Every book has its difficult passages"the company has three designs: orthogonal, cable backplane and cross-rack, each different, to be disclosed separately over the coming quarters. One thing is clear: "a lot of switches are shipping, and a lot of them are liquid-cooled." Karl Ackerman (BNP Paribas Research): Asked whether the 2027 guidance presupposes approval/financing of new power infrastructure? Does it require aggressively drawing down deferred revenue to achieve? Breithaupt: "We do not guide based on deferred revenue." The guidance is based on two quarters of visibility, conversations with our largest customers, and the run rates of the cloud and enterprise businesses. "But we see very strong demand signals." Ullal: Added explicitly: "While we don't guide deferred revenue, 2027 will, like 2026, be full of lots of new products and new use cases." Jin Marcote (Deutsche Bank Aktiengesellschaft Research): Asked how gross margin for SU-144 will compare with the existing switch portfolio once it reaches volume production in 2028? How many customers have committed? Ullal: "Since our IPO, our long-term gross margin has been 60%65%. Now we have narrowed the range to 62%64%. We don't see any change." Daniel Zhu (Bernstein): Asked about customer concentration in scale-acrossis it mainly cloud giants doing synchronous training, or also telecom operators and smaller customers, inference use cases? Ullal: Both: colos and smaller use cases will adopt smaller-form-factor platforms such as the 7280; large customers tend toward super-scale AI Spine across data centers, requiring 576-port density, terabit-scale and multi-gigawatt deployments. "Revenue concentration skews toward the 7800, while geographic concentration of multi-site deployments skews toward the 7280." Sean O'Loughlin (TD Cowen): Asked about the cycle position of front-end networkingmarket attention is all on back-end at 800G and above; what stage is traditional front-end compute at now? Ullal: "It's like a sine wave." When customers focus on the back end, front-end investment decreases, and vice versa; over six quarters to two years the two offset each other, but a given year may tilt to one sideone 10% customer once concentrated a several-hundred-gig front-end refresh in a single year, and it may reverse the following year. "Overall, we are seeing more back-end and AI networking now, and the front end will follow later." Antoine Chkaiban (New Street Research): Asked about AVA commercializationgiven several recent