楷登电子 (CDNS) 2026财年第三季度业绩电话会：营收增长22%，上调业绩指引lg...

核心要点

• 楷登电子（Cadence Design Systems，纳斯达克代码：CDNS）公布2026财年第三季度营收为16.33亿美元，同比增长22%，各产品线均实现两位数增长。
• 在手订单（Backlog）达到创纪录的82亿美元。管理层将增长归因于强劲的预订额、更广泛的客户参与以及增售机会，尽管2026年相对而言属于续约较少的年份。
• 核心EDA营收增长20%，IP营收增长24%，系统设计与分析营收增长30%。经常性营收同比增长18%。
• Non-GAAP营业利润率为44.1%，Non-GAAP每股收益（EPS）达2.11美元。GAAP营业利润率和每股收益分别为26.2%和1.05美元。
• 楷登电子将2026财年营收预期上调至63.87亿至64.17亿美元，并将Non-GAAP每股收益指引上调至8.37至8.43美元。按中点计算，管理层预计营收增长21%，Non-GAAP营业利润率为44.7%，每股收益为8.40美元。
• 管理层重点强调了其智能体AI（Agentic AI）产品组合加速普及的情况，包括100多个活跃客户项目，以及推出了针对芯片和系统设计的领域专用AI模型组合Maestro。

关键财务数据

楷登电子董事会还为公司的股票回购计划额外批准了25亿美元额度。

业务与经营业绩

智能体AI与Maestro

楷登电子推出了专为芯片和系统开发设计的领域专用AI模型组合Maestro。该平台集成了专用模型、模型路由以及与楷登电子底层EDA和系统设计工具连接的强化学习循环。

管理层表示，初步结果显示，与通用基础模型相比，其质量提升高达30%，Token效率提升高达5倍。客户可以在云端或本地部署Maestro，并可利用专有数据对模型进行进一步的后训练。

楷登电子报告称，其AI超级智能体拥有超过100个活跃客户项目。与仅由基础模型生成的代码相比，其全新的RTL生成智能体平均实现了24%的面积削减和18%的功耗降低。

联发科（MediaTek）在其生产数字验证流程中部署了ChipStack，使验证覆盖率提高了50%至70%，工程工作量减少了1.4至2倍。瑞萨电子（Renesas）报告称InnoStack带来了34倍的生产力提升，而索喜科技（Socionext）预计生产力提升高达5倍。诺基亚（Nokia）正在Millennium平台上部署AuraStack，并在选定的设计与分析工作流中实现了5倍的加速。

核心EDA

核心EDA营收同比增长20%。受对集成综合、实现和签核（signoff）工作流需求的推动，楷登电子在本季度获得了16个全新的全流程数字设计新订单。

随着人工智能和高性能计算客户增加对硬件辅助验证的使用，Palladium和Protium再次创下季度新高。管理层表示第四季度销售管道依然强劲，预计硬件业务将再创历史新高，同时提醒称季度间的交付可能会不均衡。

Spectre FX获得了更广泛的采用，包括替换了一家基础设施芯片提供商的竞争对手产品，该客户有多个2纳米项目选择了该平台。

知识产权（IP）

IP营收同比增长24%。楷登电子指出，随着定制芯片、先进制程和基于Chiplet（小芯片）架构的扩张，超大规模云服务商和半导体客户对接口、内存及基础IP的需求有所增长。

一家领先的航天技术公司为其定制芯片项目选择了楷登电子IP，这创造了开展更广泛服务和芯片合作的潜力。Tensilica创下了季度营收新高，楷登电子还宣布与亚德诺半导体（Analog Devices）就下一代音频应用展开联合开发合作。

系统设计与分析

系统设计与分析营收同比增长30%。Allegro在一家领先电动汽车客户处实现了竞品替换，该客户选择楷登电子的集成PCB设计和多物理场分析平台作为其首选解决方案。

Switch选择了Millennium和Reality数字孪生解决方案，以支持下一代AI工厂的设计与仿真。管理层还表示，收购买取的Hexagon设计与工程业务的整合进展顺利。

管理层业绩指引

在更新后的全年预期的中点，管理层预计营收增长21%，Non-GAAP营业利润率为44.7%，Non-GAAP每股收益为8.40美元。楷登电子还预计将把2026财年自由现金流的50%以上用于股票回购。

该预期假设现有的出口管制法规在本年度剩余时间里保持基本稳定。

风险与关注事项

• 楷登电子的业绩前景取决于出口管制法规在年底前保持基本稳定。
• 根据产品交付和客户安装进度，硬件营收在季度之间可能会发生波动。
• 硬件库存增加反映了公司为完成大量积压订单所做的努力，不过管理层指出，不存在影响交付的产能或供应链限制。
• Hexagon设计与工程业务在2026年贡献的营业利润有限，拉低了公司报告的边际增量利润率。管理层预计整合措施将在2027年提高其贡献。
• 在手订单和预订额提供了营收可见性，但确认的时间取决于合同结构和产品交付进度。

分析师问答环节亮点

Maestro的差异化优势与经济效益：管理层将Maestro描述为一种下一代智能体AI（Agentic AI）架构，它融合了领域专用模型、智能模型路由以及与楷登电子工具的紧密集成。楷登电子计划支持自身模型以及领先的前沿模型和开源权重模型，允许客户选择云端或本地部署。该公司未提供具体的营收目标，但表示工作流使用量的增加应为楷登电子创造更大价值，同时保持以软件为主导的经济效益。

经常性营收加速增长：管理层将18%的经常性营收增长归因于智能体AI的普及、底层EDA软件使用量的增加以及良好的客户拓展。楷登电子强调，AI智能体会不断调用其设计引擎，这可能会增加对核心工具的消耗。

硬件需求：管理层表示，随着芯片和系统复杂性的提升，硬件辅助验证变得越来越重要。楷登电子预计需求将受益于更大规模的设计、3D-IC架构以及共同验证多个小芯片（Chiplet）的需求。

经营杠杆：该公司表示，预计2026财年营收将增加约11亿美元，Non-GAAP营业利润将增加约5亿美元。管理层指出，如果不计入收购Hexagon设计与工程业务的影响，增量利润率与楷登电子通常54%–55%的区间相符。

EDA市场地位：管理层指出，公司在数字实现领域的份额有所提升，涵盖综合、实现和签核的全套数字流程的普及度不断增长。管理层还强调了模拟设计和验证产品的优势，包括Palladium、Jasper、Verisium和ChipStack。

定制芯片的扩张：楷登电子预计，更易于使用的AI设计工具将降低非传统芯片开发者的门槛。管理层提到超大规模云服务商、汽车、航天、金融和机器人公司的活动日益活跃，这些领域都能为EDA软件、IP、硬件和设计服务带来需求。

业绩电话会议完整文字记录

---

完整财报电话会议逐字稿

管理层陈述

Operator

Good afternoon. My name is Jade, and I will be your conference operator today. At this time, I would like to welcome everyone to the Cadence Third Quarter 2026 Earnings Conference Call. [Operator Instructions]

I will now turn the call over to Richard Gu, Vice President of Investor Relations for Cadence. Please go ahead.

Richard Gu

Thank you, operator. I'd like to welcome everyone to our third quarter of 2026 earnings conference call. I'm joined today by Anirudh Devgan, President and Chief Executive Officer; and John Wall, Senior Vice President and Chief Financial Officer. The webcast of this call and a copy of today's prepared remarks will be available on our website, cadence.com. Today's discussion will contain forward-looking statements, including our outlook on future business and operating results. Due to risks and uncertainties, actual results may differ materially from those projected or implied in today's discussion. For information on factors that could cause actual results to differ, please refer to our SEC filings, including our most recent Forms 10-K and 10-Q, CFO commentary in today's earnings release. All forward-looking statements during this call are based on estimates and information available to us as of today. and we disclaim any obligation to update them.

In addition, all financial measures discussed on this call are non-GAAP, unless otherwise specified. The non-GAAP measures should not be considered in isolation from or as a substitute for GAAP results. Reconciliations of GAAP to non-GAAP measures are included in today's earnings release. For the Q&A session today, I would ask that you observe a limit of one question only. If time permits, you can re-queue with additional questions.

Now I'll turn the call over to Anirudh.

Anirudh Devgan

Thank you, Richard. Good afternoon, everyone, and thank you for joining us today. I'm pleased to report that Cadence delivered another outstanding quarter with all key metrics coming in well above our guidance. Backlog reached a record of $8.2 billion, reflecting strong bookings driven by growing customer demand. Given our strong business momentum, expanding technology leadership and compelling AI growth opportunities, we are raising our revenue outlook for the year to 21% growth, and our EPS outlook by $0.30 to $8.40. John will provide more details on both our Q3 results and the updated financial outlook.

Agentic AI represents a transformational market opportunity, and Cadence is exceptionally well positioned to capitalize on it. Our differentiated three-layer framework is anchored by our physically accurate simulation and optimization engines in the critical middle layer, tightly integrated with AI agents and orchestration at the top and accelerated compute and proprietary data at the base. Agentic AI continues to evolve at a remarkable pace. With our AI super agents gaining strong traction, we are now addressing the long-standing customer demand for domain specialized models. We believe the future of Agentic-AI-based design is increasingly domain specialized models combined with intelligent model routing for performance, and reinforcement learning loops with underlying tools for recursive self-improvement.

Today, I'm excited to introduce my stroke, a portfolio of domain specialized AI models, purpose build for chip and system design. By controlling all three layers, our AI super agents, Maestro models and proven EDA and SDA tools, we are uniquely positioned to deliver optimal performance, efficiency and throughput results for our customers. Initial results are very compelling, demonstrating up to 30% higher quality and up to 5x greater token efficiency versus general purpose foundation model. Moreover, the Maestro models can be further post trained on customers' proprietary data to deliver even better outcomes. Our model router chooses among our models and other leading frontier and open wide models as needed for optimal results. Customers can opt to deploy Maestro models in the cloud or in their on-prem environment. Maestro is gaining strong early customer tractions with several deployments underway.

Momentum of our AI super agents continues to accelerate with compelling early results and more than 100 active customer engagements. We added a new agent to chip stack for RTL generation and early PP optimization. That created RTL delivering an average of 24% area and 18% power reduction versus pure foundation model generated code. Importantly, this demonstrates the power of combining our AI super agents with AI models and our underlying EDA engines to achieve outcomes that foundation models alone cannot deliver.

Honda is using the RTL generation agent and AI-powered automation to improve productivity from specification to RTL development. In Q3, MediaTek deployed chip stack into its production digital verification flow, achieving 50% to 70% improvement in verification coverage and 1.4 to 2x reduction in engineering efforts with deployment now scaling across design teams. In analog and custom design. ViraStack saw 4x increase in customer engagement with deployments in various production designs at multiple customers. The combination of agentive workflow, embedded skills and our Virtuoso platform is unlocking significant new productivity opportunities in a traditionally difficult to automate domain. We are seeing strong customer interest in InnoStack, our digital AI super agent ad customers push to improve PPA, accelerated design closure and boost productivity.

In Q3, the Renesas MCU team successfully deployed InnoStack across a wide range of engineering costs, achieving a 34x productivity improvement with plans to expand its usage to layout automation. Also in Q3, Samsung Foundry successfully used InnoStack to dramatically improve engineering efficiency and accelerate PPA exploration, reshaping its approach to technology migration and BTCO optimization.

Socionext is deploying InnoStack in its production design environment. And I expect its combination with our advanced physical design tools to deliver up to a 5x overall productivity gains while maintaining design quality. In PCB and 3DIC design, the early response to our AuraStack AI Super Agent has been overwhelmingly positive. Nokia is deploying AuraStack on our Millennium supercomputer to advance electronic system design at scale. By unifying Agentic AI, our physics-based solvers and accelerated compute in a secure on-prem platform, Nokia is achieving a 5x acceleration in key design and analysis workflows.

Moving on to our businesses. We are delighted that all product groups delivered outstanding results with year-over-year revenue growth of 20% in core EDA, 24% in IP and 30% in SD&A. Accelerating design activity continues to drive strong demand for our core EDA, SDA and IP products. At the same time, we are seeing amazing customer momentum for our Agentic AI portfolio, which is also fueling increased demand for our underlying products. We are pleased that recurring revenue sustained its strong momentum, growing 18% year-over-year. I'm particularly proud of the significant wide-ranging agreement we have signed with top customers. We furthered our long-standing partnership with a global marquee systems customer through a significant wide-ranging expansion of our core EDA, hardware and SDA solutions.

In Q3, Cadence and a top telco deepened their long-term strategic partnership with a broad expansion across our AI offering, EDA software, hardware and SDA solutions, establishing Cadence as their primary design partner. Core EDA revenue growth was driven by broad-based strength across the portfolio. Our digital and sign-up business continue to expand, achieving 16 new full flow wins during the quarter. As agentic flows increasingly require full vertical workflows, there is growing demand for our synthesis and sign-off solutions. Our sign-up product saw growing adoption and our verification software product had strong growth in big digital and memory designs with increasing usage at major hyperscaler and memory companies. As AI and HPC designs continue to grow in scale and complexity, hardware-assisted verification has become an increasingly strategic capability for our customers. This is driving sustained unprecedented demand for our Palladium and Protium emulation platforms, which delivered yet another record quarter. We added several new logos and saw a meaningful expansion with marquee customers across a wide range of verticals.

Our hardware cloud offering showed strong growth with a leading semiconductor company adopting more hardware cloud services to accelerate new product development. In analog and custom design, Spectre FX continues to gain momentum, highlighted by major competitive displacement at a marquee infrastructure silicon provider with multiple new 2-nanometer projects adopting our platform. Our IP business delivered another outstanding quarter, driven by deepening engagements with leading hyperscalers and semi customers. We continue to benefit from strong AI-driven tailwinds, fueling accelerating AI infrastructure investments, growing adoption of custom silicon and increasing demand for high-performance IP. Along with the transition to advanced nodes and chiplet-based architectures, these AI-driven trends are creating a strong full for our differentiated STAR IP portfolio across interface, memory and foundation IP. We are well positioned for sustained growth as we continue to deliver superior power performance and area across our expanding advanced IP portfolio.

A leading space technology company, selected our IP portfolio for their custom chip program, paving the way for a broader services and silicon engagement. Tensilica delivered a record revenue quarter and we announced a joint development with analog devices across next-generation audio applications.

Our system design and analysis business delivered another strong quarter, driven by growing adoption of our integrated solutions and proliferation of our expanded electronics and physical systems intelligence portfolio. Allegro momentum continued, including a competitive displacement at a leading EV customer that chose Cadence's integrated PCB design and multiphysics analysis platform as its primary solution. Our Millennium platform had strong adoption driven by deployment of our GPU-accelerated electronic and mechanical systems suite, along with our reality digital twin solution.

During the quarter, Switch, a leader in sustainable high-density AI infrastructure selected our Millennium and reality digital twin solutions to advance the design and simulation of its next-generation AI factories. The MSE business is the integral part of our physical AI strategy, and its integration is progressing well. MSC Adam our industry-leading multibody dynamic solution helped secure a competitive displacement at a leading power company, along with adoption of other cadence products.

In summary, Cadence delivered exceptional Q3, and I'm delighted with the accelerating momentum of our business. We are executing extremely well in a vibrant AI market, have our strongest competitive position ever and are extending our leadership in AI. Our highly recurring and ratable business model, combined with growing customer R&D investments is positioning us very well for continued strength in this dynamic AI environment.

With that, I will turn it over to John to provide more details on our Q3 results and our updated 2026 outlook.

John Wall

Thanks, Anirudh, and good afternoon, everyone. Cadence delivered strong third quarter results driven by robust demand across our businesses and growing momentum from AI-driven designs. Growth is accelerating and broadening Q3 revenue grew 22% year-over-year with strong double-digit growth across all products. We exceeded our guidance on every key financial metric and ended the quarter with record backlog of $8.2 billion.

Here are some of the financial highlights from the third quarter, starting with the P&L. Total revenue was $1.633 billion. GAAP operating margin was 26.2%. Non-GAAP operating margin was 44.1%. GAAP EPS was $1.05. Non-GAAP EPS was $2.11.

Next, turning to the balance sheet and cash flow. Our cash balance was $1.452 billion, while the principal value of debt outstanding was $2.650 billion. Operating cash flow was $424 million. DSOs were 66 days, and we used $400 million to repurchase Cadence shares during the quarter, reflecting our confidence in the business and our commitment to returning capital to shareholders. Underscoring that commitment, our Board has authorized an additional $2.5 billion for our share repurchase program.

Before I provide our updated outlook, I'd like to highlight that it contains the usual assumption that export control regulations that exist today remains substantially similar for the remainder of the year. For our updated outlook for 2026, we now expect revenue in the range of $6.387 billion to $6.417 million. GAAP operating margin in the range of $28.2 to 29.2%. Non-GAAP operating margin in the range of 44.2% to 45.2%. GAAP EPS in the range of $4.84 to $4.90. Non-GAAP EPS in the range of $8.37 to $8.43. Operating cash flow of approximately $2 billion, we expect to use more than 50% of our free cash flow to repurchase Cadence shares in 2026.

For Q4, we expect revenue in the range of $1.695 billion to $1.725 billion. GAAP operating margin in the range of 30% to 31%. Non-GAAP operating margin in the range of 44% to 45%. GAAP EPS in the range of $1.23 to $1.29, and and non-GAAP EPS in the range of $2.19 to $2.25. And as usual, we published a CFO commentary document on our Investor Relations website which includes our outlook for additional items as well as further analysis and GAAP to non-GAAP reconciliations.

Overall, I'm pleased with our Q3 results, especially our recurring revenue growth of 18% year-over-year. At the midpoint of our updated 2026 outlook, we now expect revenue growth of 21%, operating margin of 44.7%, and EPS of $8.40. As always, I'd like to close by thanking our customers, partners and our employees for their continued support.

And with that, operator, we will now take questions.

Operator

[Operator Instructions] Your first question comes from the line of Jason Celino from KeyBanc Capital Markets.

分析师问答

Jason Celino

Anirudh, really nice to see the Maestro in announcement. You know I'm not an EDA engineer, I'm not an AI scientist, so maybe can you just explain what this is a little bit better. It sounds like it's part model router, but also maybe domain-specific models. So is Cadence developing these main specific models? Maybe just elaborate, and then how the monetization of this new product will work?

Anirudh Devgan

Yes. Hi, Jason, and thank you for the question. Yes, there's a lot in that release. So let me try to unpack it a little bit. First of all, we are very pleased with the momentum of our agent stack and all the four super agents. And you can see there are like more than 100 engagements now. So as we work with customers over the last 6 to 9 months, several trends became pretty clear, okay? And to me, we are launching Maestro as what I would call the next generation of agents. And there are three big factors in the next generation of agents. So one is, like what has happened in the last one year is that, of course, we work with all the Frontier labs from the very beginning, whether they're doing chip design themselves, or we are using their LLMs. But over the last one year, there is this emergence of very, very good open rate models, okay? And I always wanted to do like domain specialized models, but now it's finally possible with open weight models. And it's actually very cost effective to post train these models, to deliver like domains intelligence, which is remarkable. We're getting up to 30% better quality by using our tools and data to post train these models. And also, you don't need like -- it's a portfolio of model. So all the way to like small decisions LLMs to small and medium LLM all the way up to fronter class open wage models post trained on our data and intelligence. So that's the first part of Maestro. So it gives really customization for chip and system design and gives much, much better results. It's actually quite remarkable how much these models can be improved by post training, okay. Now the second part of that is in our agents and when they do these tasks have a variety of tasks, right? So we don't need only one size or type of model to do the task. So the second part is model routing. Just like you don't need to use an airplane to go to your local grocery store, right? So similarly with agents, you have a variety of tasks that we need to do from simple task to like deep thinking and reasoning tasks. So the moderator will efficiently knowing what the task is from our super agents use the right model. And that can easily give like up to 5x improvement in token efficiency. So either much lower cost or much more work for the same cost. And then the third thing is which I always talked about, this reinforcement learning loop with the underlying tools. And we have great examples we talked about our new RTL agent that one of the customers publicly endorsed Honda gave 24% area and 18% power because the integration with the underlying tools, and how we call them, and continuous self-improvement whether it's skills or models gives us a unique advantage. So if you combine these three things together, domain specialized models -- domain specialized family of models, plus model routing plus deep integration with our tools and data, it gives a unique differentiation that nobody else can do. And it also allows us to control all three layers of the cake and innovate in all three layers of the cake. So I'm actually super excited. And I think this is what's going to happen in all industries as these LLMs mature, this domain specialized intelligence model routing and integration with underlying engines.

Operator

Your next question comes from the line of Jim Schneider from Goldman Sachs.

James Schneider

I was wondering if you could maybe give us an update on the monetization of the Agentic AI solutions you're seeing. I believe last quarter, you talked about recurring revenue growth of 20%. This quarter was 18%. Maybe talk about what the difference there and the recurring was? And maybe underlying that, whether there was any acceleration in the Agentic part of that recurring revenue.

Anirudh Devgan

Yes. Thanks for the question. I'll start, and then John can comment. The -- I'm very proud of the recurring growth. And I mean, this quarter, there were some specific things. Like if you compare it to Q3 of last year, there were some specific things with China revenue moving around, which is not there this year. But in general, recurring revenue is doing phenomenally well. And if you step back and compare last year, recurring revenue was 11%, and it's 18% for this quarter, and it was very good last quarter. So I expect recurring revenue to do very well in 2026. And this is a result of, of course, our Agentic AI solutions, but also Agentic AI solutions driving more usage of our base tools, which, as we always said, is going to happen. So that's what we see. A lot of engagement is [Audio Agentic and a lot of usage of our base tools. So that provides a lot of growth for recurring and then you add that with our ongoing strength in IP and hardware, and you get a very, very good set of results.

John Wall

Yes, I'd agree with that, Anirudh. I mean it basically reflects strong demand across our software portfolio and healthy customer expansions and continued strength in core EDA. Yes, I think that's the way to look at it in terms of strong high teen growth in recurring revenue. We're very pleased with this.

Operator

Your next question comes from the line of Vivek Arya from Bank of America Securities.

Vivek Arya

Anirudhn, we recently heard three things from your competitors that they can grow sales kind of at a mid-teens pace next year and then compound it that they can achieve 50% operating margins? And then also, interestingly, there is a royalty-based model that can help add nearly $1 billion to their top line over the next several years. So I realize, right, this is not the call to give a multiyear growth outlook, but I was just hoping to get your perspective on how you envision sort of this growth and royalty component of Cadence, and how well is Cadence is positioned to maintain or expand your market share and your leading operating margins over time?

Anirudh Devgan

Yes. Thanks for the question, Vivek. So overall, I'm very pleased with our growth and not just growth, but profitability and margin. And we always talk about like a rule of 40 metric or revenue growth plus margin. So I think we will end the year, if you look at our guide, about 21% revenue growth and slightly less than 45% margin. So we will end up north of 65% in 2026, okay? That's the important thing to point out Okay. And we still have one more quarter to go. So we'll see -- it's an important quarter for us. Fourth quarter is always important. So we'll see how that finishes, and we can update you like we usually do in February for the next year. But overall, the environment, I think, is very positive for us. As you know, the AI build-out continues the amount of design activity in all aspects of AI and even the traditional companies and analog mixed-signal companies, digital company, memory companies, system companies, so the amount of design activity is the strongest that I have seen. And our -- we're gaining share in almost all segments of our business, whether it's IP or hardware or EDA, Agentic AI. So our product portfolio, I think, is the strongest it has ever been. And then we have this huge opportunity with Agentic AI. And I think we are redoing how agent will be defined for all industries with our own domain specialized models and the flexibility to continue to work with all the frontier models. So if you look at all these three things, I feel Cadence is in a very, very strong position. Now in terms of your specific question on royalty, we are always looking at different kind of business models. So in IT, we always had royalty for a while in Tensilica, and now we have royalty in our design IP business as well. But to me, that's just one vector. We're always looking at how to best serve -- give value to our customers and then get, of course, value and return.

Operator

Your next question comes from the line of Siti Panigrahi from Mizuho.

Sitikantha Panigrahi

I just want to ask on hardware I see the inventory again grew this quarter, I'm wondering hardware is a strong quarter. How sustainable hardware growth for heading into next year 2027. Any color would be helpful there. And a quick clarification, John. There was a spike in G&A expense in Q3. Was there anything onetime there?

John Wall

Yes. Sure, Siti. Great. Look, I'll take the hardware question first. The hardwood continues to perform exceptionally well. I mean we delivered another record quarter, supported by strong demand from AI and HPC customers. And we're also seeing growing interest in cloud-based hardware access. The hardware-assisted verification is becoming increasingly important because customers are designing systems of unprecedented complexity. They need the capacity to verify those designs before committing to silicon. We have a strong pipeline entering Q4. We continue to expect another record year for hardware as always, harder could be lumpy from quarter to quarter depending on product deliveries and customer stellate schedules. But the underlying demand environment remains very strong, very healthy, and we're very, very encouraged by the pipeline of customer engagement. In terms of inventory, the increase in hardware inventories to support customer orders that we have a significant backlog that we're trying to work our way through, but there are no capacity or supply chain constraints in terms of addressing that. We're just building inventory to take those numbers. I think you asked about that's in part of the quarterly movement this quarter reflects the timing of annual cherishable contributions, which occurred in Q3 this year compared to Q4 last year. That would extrapolate that timing difference to the underlying G&A run rate. So there's a slight impact on G&A in Q3 because of that -- those cherishable contributions.

Anirudh Devgan

Yes. And Siti, just to add on the hardware piece. I mean Palladium and Protium performing exceptionally well. Not just for AI. And of course, we are working with all the major AI companies, but also any other kind of chip design right now, hardware became a indispensable, almost indispensable way to design these things. Now one thing is always like will this trend continue? And we'll see, but there's one megatrend, okay? So hardware is -- you probably know this already, but just to highlight it even further. Hardware is normally sold or consumed based on design size. So if the design size is bigger, then you need more and more hardware. So actually, I was recently just an example of what is going to happen to design, I mean the design sizes are already big, but what's going to happen to design size in the future. So just to give you an example, I was that IMEG recently, I gave a talk at IMEG, which is a European research consortium. And TSMC will also gave a talk same session. So TSMC showed their process road map and said in the next 5 years, the chip size will go up by 48x. Now typically, this would happen to be like 5 to 10x in the old days, but I think it's going to be 48x because you have to add regular technology scaling plus 3D IC. So all these 3D IC packaging wafer scale integration. So that's going to even further create more and more demand for hardware and die-to-die IP and things like that. But -- so now our customers are not only verifying 1 chip or 1 chiplet, they're verifying those chiplet working with other chiplets and the size of the chips is going up and then a amount of chiplits are going up. So I think this trend, 1 criticality of hardware-based design and verification combined with the growth in size is going to continue for some time.

Operator

Your next question comes from the line of Joe Quatrochi from Wells Fargo.

Joseph Quatrochi

I was curious if you could just kind of talk about the change in the '26 outlook, particularly from an operating profit perspective because I think when you flow through just kind of the incremental margin on the upside of revenue, it's quite significant relative to maybe your kind of targeted over 50%. So maybe just any help there in terms of understanding what's going on there.

John Wall

Sure. Yes. Look, we continue to be very pleased with the underlying profitability of the business, despite integrating Hexagon's DD business and making targeted investments in our strategic growth opportunities, we're now expecting approximately 44.7% non-GAAP operating margin for 2026, which is slightly above last year. But if you look at the revenue growth this year over last year, it's about $1.1 billion. Non-GAAP operating profit is growing about $500 million. And would you consider the Hexagon's DEV being part of that. And that's not really contributing much to the way of operating profit at all this year. But -- and I think we said at the start of the year, we're expecting Hexagon to be about $160 million for the year, is doing slightly better than that, but it's pretty much on target. But I mean, if you back that out, we'd be where you typically expect us to be, which is kind of mid-teens for revenue growth of high teens for EPS growth in our usual kind of 54%, 55% incremental margin rate. So the Hexagon DD is kind of impacting the incremental margin for this year. But we'd expect that to help us next year because the integration go really well. And we're taking steps now to improve things for next year in the same way as we did with paid the last couple of years.

Operator

Your next question comes from the line of Charles Shi from Needham & Company.

Yu Shi

Glad to hear about the Metro model portfolio just announced today. In light of your competitor who announced the GPG synopsis, can you tell us what your thoughts are regarding collaborating maybe with the Frontier labs for a specialized model, given that Maestro based on what you just described, it sounds like it's a collection of open models post train by cadence. And I think [indiscernible] customers also have choices between your model portfolio, and let's say GPT synopsis, what do you tell your customers when they have both choices and why they should take Cadence?

Anirudh Devgan

Yes. Thanks for the question, Charles. So first of all, I just want to point out that we have worked with open AI from the very beginning of when they got started. So in terms of chip design, and we have all kinds of collaborations with open AI. And same thing with Entropic and Gemini, the three leading frontier model. So we always discuss all kinds of collaborations with them. Now in terms of what we want to make sure as -- in Maestro, the key thing is Maestro is not just our own models, but it still has maximum flexibility to use the right model for the right task. So let's say, if certain for a certain percentage of task, a certain frontier model is better, we will, of course, use that to give the best performance to our customers. But the key thing in our solution is that we want to control the three layers of the cake, and we want to control the interaction with our customers. We want to be the primary interface to our customers, which is what our customers want. And in return, our agents will provide them a lot of value, and we'll use the right LLM for the task, okay? But the critical thing is as we became the primary interface, and we guarantee the confidentiality of our data and their data. So the other thing we will allow with Maestro is if a customer wants to post train Maestro further, and this is another request from our customers, we will allow that. If the customers want to have it on-prem, we will allow that. If the customer wants to host on the cloud, we will allow that, okay? So it gives maximum flexibility and maximum performance by combining all 3 layers and only Cadence can provide this. And that will give without any doubt, the best results for our customers.

Operator

Your next question comes from the line of Harlan Sur from JP Morgan.

Harlan Sur

Last quarter, you mentioned the pickup in the more traditional analog power microcontroller segments of the market, right, targeting industrial, automotive, consumer IoT, that's the positive cyclical dynamics we're seeing now as excess customer inventories have bottomed in these end markets. Are you continuing to see the positive demand trends in this segment? And do these same companies are now very focused on embedding next-generation physical AI-based capabilities into their chip designs, right your MCU customers, for example, are adding NPU capabilities to their trip designs, our analog customers are designing more sensor-related capabilities into their chips more ADD/DDA converters on top DSP processing. Is this also translating into a better book of business for the Cadence team?

Anirudh Devgan

Yes, Harlan. I guess a great point. So I think the momentum is very strong in analog and mixed signal, and you can look at their results as well, driven by physical AI and overall recovery of the market. And of course, memory, which we work with all the big memory companies is having, of course, amazing improvement from a year or two ago. And I think what it shows up is, first of all, we are very, very strong in mix-signal companies. So because we have the most complete portfolio with both analog and digital, so these are designs like you mentioned, are more and more mixed signals. So that plays to Cadence's trend. And also ViraStack. I'm really pleased by the momentum in ViraStack because as analog customers, whether they are mixed signal companies or these big AI companies also have huge team. I don't know if you realized that some of these big AI companies designing their own IP and all have huge analyst teams. So there is a massive demand for analog productivity, which we could not provide earlier. I mean we always wanted to. But with Agentic AI, we can do a lot more in analog productivity. And you can see that in some of the customer codes we have seen and the amount of engagements we have with ViraStack. So overall, I'm pretty pleased not just with the environment of analog mixed signal, but also the new capabilities we are launching with ViraStack.

Operator

Your next question comes from the line of Joe Vruwink from Baird.

Joe Vruwink, a reminder to unmute your device. Your line is open.

Your next question comes from the line of Lee Simpson from Morgan Stanley.

Lee Simpson

Anirudh, it does seem as though you and your peer are pursuing very distinct routes to monetizing agent in chip design. I mean, I thought you answered one of the previous questions quite well. It really does bring out that you're looking to have your own control on the modeling your deployment flexibility comes through. So you can see that adoption, as you see it versus a frontier model-centric partnership is something you think will gain quite good adoption. But is there something in the performance that we need to better appreciate or even in the delivery costs that really led you to go down this route in strategy? And maybe just a follow-up question. You did mention that a leading space company was adopting some of your IP. Just wonder if there's anything specific to call out there. Are they taking just typical IP blocks, or does it come with any broader IP services attached?

Anirudh Devgan

Yes. Let me answer the second part first, I come back to the first one. I mean, in general, like we have mentioned that now 45% of our business is coming from the system companies, the nontraditional versus the traditional semi, so these are new customers. And what is also encouraging to see is that it's not just big AI, hyperscalers whether doing design. But we are seeing a lot of other system companies like the one we highlighted like a space company even companies in finance doing their own chips. Okay, this is -- as we make chip design easier and easier, this is always the promise that more and more companies we'll do that. And when these new companies do design, we naturally have IP opportunities but also hardware and also naturally design services because they're getting started. And we have one of the best design services capabilities as we are the best at using our own products. So I think this is some space we have to watch, but I'm pretty encouraged by more system companies doing chip design. Now coming back to the to the AI solution, we want to have control and we want to give flexibility to our customers. So this is the right way to look at it. So -- and there is a lot of value in the model itself, and it is actually pretty cost economical to post train a frontier class model. Okay, this could not have happened like 2 years ago. But now it's actually pretty manageable cost to -- and we get pretty remarkable results, like I mentioned, by very incremental cost to post train a frontier model. And of course, that's an important thing to have in the bag along with a commercially available frontier model. It's just using the best solution of the two. And then in our own genic development, right? We have our own IP team, a lot of people using our -- of course, using agent stack and all the solutions. So the benefits of costs apply to us as well. So even though there is some incremental expense to train them, there's also incremental saving in our own deployment of AI. So overall, I think it should not affect any financial model of cadence it only gets stronger as we add like a new kind of revenue capability and all the margin progression and all should though as we have planned for a while. But what Maestro does is this flexibility, but also gives customization ability, it gives control and then it gives cost savings. So this is best of all words.

John Wall

Yes, Lee, if I could add there. I think that that's a very, very important point that Andrew makes in terms of the flexibility. We really want customers to have choice and control over the models they use while benefiting from cadences, debate expertise and deep integration with our design tools. I mean our strategy remains centered on delivering high-value engineering software rather than building a large general purpose AI infrastructure business. We're focused on improving the efficiency and the value of agentive workflows and maintaining attractive software economics as customer adoption scales, but we think the flexibility works in our model. Our model allows customers to use their old LLM infrastructure bear their associated bottle compute cost if they want to, a Cadence, the underlying design and orogenic orchestrate capabilities. The economics of particular deployment options could vary, of course, but the principle is that greater customer usage should create greater value for Cadence. So I think that's where we differentiate ourselves.

Operator

Your next question comes from the line of Joe Vruwink from Baird.

Joseph Vruwink

Can you hear me?

John Wall

Yes.

Anirudh Devgan

Yes, Joe, please go ahead.

Joseph Vruwink

I want to go back to ViraStack and InnoStack. I think it's that the analog and back-end areas haven't seen the AI proliferation yet like front-end design has. Are you seeing this change is specialized models and post training, something that helps better address these domains? And if all of this is starting to happen, is there a commensurately higher revenue implication for Cadence relative to even what has been seen to date?

Anirudh Devgan

Joe, I mean, first of all, it there is we started with chip stack, but then we have launched like 3 other major super agents, as you know, ViraStack for analog, InnoStack for back end, and AuraStack, the latest one, for package and PCB. And I have to say the demand is actually phenomenal in mind is phenomenal, and especially because these chip stack is doing great, but also these other agents are even more integrated with our base tools. So InnoStack, if you look at our endorsement, typically, when we work with customers to get a public endorsement is only happens like 1 out of 10 customers because not everybody wants to talk about what they're doing, but we have maybe 10 endorsements this quarter. So that tells you the extent of the deployment of all these agents. And if you look at InnoStack, Renesas talking about like 34x productivity improvement, Samsung Foundry, see, there are other use cases that can happen, not just physical design. One big use case is design technology co-optimization with the foundry. So the InnoStack a critical agent for that. Socionext talked about 5x productivity improvement. [ Rapido ] talked about 2x improvement. So there's a multiple x improvement happening with InnoStack. ViraStack take already said that demand is off the charts. And AuraStack, this the latest area where innovation hasn't happened. So Nokia showed 5x productivity improvement with AuraStack. So ChipStack is probably a year now in deployment, and these are maybe 6 to 9 months, but there is acceleration deployment of these -- the newer, but they are even more tightly integrated with the base tools. And then Maestro adds further differentiation on top of these capabilities. So I feel very, very confident. And what's going to happen is we have a lot of R&D investment in our traditional tools in the middle layer, but now we are putting significant differentiation and investment in the top layer. We are defining, I believe, what will be the identic flow for the future in all industries. So as we make the Agentic, the top layer, and more and more complicated, more and more integrated, more and more differentiated, is very difficult to compete with Cadence in the Agentic AI, whether you're a start-up or if you are trying to write -- a customer tried to write some agents, but the capabilities with motor routing with domain-specific models with tight integration with agent and the tools. This is something that our customers can't do themselves. So that also bodes very well for all of these customers realizing that and working more and more closely not just in the middle layer, but in the top layer with us. So I feel very good in terms of how the speed and differentiation of our Agentic offerings.

Operator

Your next question comes from the line of Jay Vleeschhouwer from Griffin Securities.

Jay Vleeschhouwer

Anirudh in your prepared remarks, you made some important references to the components of your core EDA, particularly in digital. It's demonstrable that over the last number of years, you've gained considerable share in core EDA. But I'd like to delve into what the components of that share gain have been. In other words, when we think about the various large categories that comprise core EDA, especially digital, would it be fair to say that the single largest difference in terms of the dollar and share gain you've had over the last number of years has been with Innovus in implementation and perhaps, to a lesser degree, in areas like synthesis, RTL simulation, analysis and so forth. Or perhaps you can speak about the breadth of where that incremental revenue and shares come from in core EDA to the extent that it's not really largely concentrated in banks and route.

Anirudh Devgan

Yes, Jay, great question. So yes, I mean, over the -- if you look at over 5, 7 years, we have gained so much in digital implementation, which I was becoming basically the platform of choice for advanced nodes design. And then what is particularly encouraging in the last 6 to 12 months is not just Innovus, but all the tools around Innovus. And we always -- I have talked about for 10 years now. Of course, there's integration at lower nodes. So that was going to happen anyway as we go to 3-nanometer, 2-nanometer, it's difficult to separate sign-off from implementation or synthesis from implementation. But then with this Agentic AI, there's a further kind of need for vertical flows, which are well integrated. So what I -- and I mentioned this last quarter also, I'm particularly pleased to see -- I mean, not only our tools and synthesis and sign-off are much, much better now versus a few years ago. that we are getting this real full flow deployment at not just small or medium customers, but some very, very large customers are deploying our full digital flow, and this is a great trend for Cadence. At the same time, we maintain our undisputed leadership in analog, okay? And then because of the strength of Palladium, strength of Jasper, strength of Verasium and Agentic Flow ChipStack, it is also helping our verification business and logic simulation. So overall, that's what I was saying, I feel that all the three big parts of EDA, we are in a very, very strong position and gaining momentum, not just individual tools being better and proven by large companies, but also this integration that advanced integration with the agentic solution. So we'll see how it is. But overall, I feel pretty good about it.

Operator

Your next question comes from the line of Gary Mobley from Stone X.

Gary Mobley

Your bookings in the somewhere in the ballpark of, I don't know, mid-20%, call it, 26% year-over-year. If I recall correctly, this was supposed to be a low renewal period 2026 was supposed to be. So my question is, what's driving the strong backlog, the strong bookings? Is it utilizing traditional EDA tools as a result of AI agent workflows. Is it a reflection of the strength of the semiconductor industry? Any color that you can add there would be helpful.

John Wall

Sure. Thanks, Gary. I mean, we're very pleased with the both the bookings performance and the growth in backlog, particularly given that 2026, as you mentioned, is it comparatively light new year for us. The strength reflects the underlying desire activity, expanded customer engagement and strong add-on opportunities across our portfolio. And what we're seeing is that Cadence is becoming increasingly important to customers as they look to accelerate their engineering road maps and address more complex design challenges, backlog itself provides good visibility into that. But it's important to revert that backlog and bookings number reflects the timing of revenue recognition that depends on the structure and delivery of a digital ratio stuff like that. But -- so there's always a blade type add value in those. But strong bookings in a lighter renew here are particularly encouraging because they reflect expanding customer engagements adoptively the renewal calendar. But we're very, very pleased with that booking strength.

Anirudh Devgan

So Gary, needless to say, I mean, we've done three quarters. There's still one more quarter to go, which is traditionally [indiscernible]. So we'll see where we end up.

Operator

Your next question comes from the line of Lachlan Brown from Rothschild.

Lachlan Brown

Are you seeing much evidence that new AI design tools like ChipStack lowering the barriers to intrigue for custom silicon for nontraditional designers. And I appreciate that land expand naturally takes time. But are there signs that new logos could become more of a meaningful growth driver for that TAM expansion going forward?

Anirudh Devgan

Yes, absolutely. I mean like I've talked about this for years that as we make chip design easier to do, there will be more and more people doing chip design. And if you step back, we made chip design maybe over the last 20 years, 100x easier to do because maybe 20 years ago, all these IBM Intel would do like CPU design with 500 people in like 5 years. Now you can do a CPU design 6 months with like 50 people. So that's like 100x better. And that's what led to more and more designs. And with AI, we can have and add another 5 to 10x. So it will be like 1,000x easier than probably 20, 25 years ago. And this was invariably lead to more and more design activity. And you can see that even with the hyperscalers, if you go back 1 or 2 years ago, it was not clear that the hyperscalers will do or we always get questions will they continue to do, but now with Google and Meta and Microsoft, Amazon and Tesla, everybody is doing more and more designs. Now at the same time, the traditional semiconductor companies are also doing well, of course, the big semi companies. And this is only going to accelerate, in my opinion, with -- like we talked about space companies, finance companies, physical car companies, the amount of design activity we are seeing in automotive is actually quite remarkable and robotics and things like that. And then these things, robotics also require a lot of different chips, right? The amount of chips that are there in one hand, there are like 25 chips in one hand of a humanoid robot to control all the individual motors. So I think we are -- I believe this for years, and I think you will see more and more of this. And our job is to make these things easier to use, have more -- give more flexibility to our customers, like we did with our agent solution in Maestro, so they can do more themselves.

Operator

Your final question comes from the line of Kelsey Chia Citi.

Wei Chia

Anirudh and John, congrats on great results. My question is on EDA. EDA growth has been very strong for the last three quarters, 18%, 19% year-on-year growth well above our historical growth rate. And it sounds like a majority of that growth is primarily driven by hardware and agent attribution is still quite small. Is that a fair characterization? Or could you give us a sense of customer adoption levels for identic workflow today? And as much of the sort of revenue acceleration from Agentic AI as to ahead of us?

Anirudh Devgan

Yes, thanks for the question. Yes, of course, we are very proud of the hardware growth. But as you know, we will break out recurring and nonrecurring growth. So like I was mentioning earlier, and John mentioned also, our recurring growth has improved from 11% last year to 18%. So that is not hardware. That's ex hardware because hardware will typically show up in nonrecurring growth. So that's a major improvement of our recurring growth from last year to this year. And that is a combination of Agentic AI plus the growth of underlying tools and design starts on the software side.

John Wall

Yes. Kelsey, I think based on lists to answer Anirudh's question there a few minutes ago that essentially, we believe that Agentic AI is a demand accelerator for EDA. I mean customers have always built their own workflows around Cadence tools. And we expect them to continue to develop their own agents as well. Our strategy accommodates that. What's exciting because that agents could perform much more designed exploration than an individual engineer, calling our underlying engines repeatedly, and that creates a potential for significantly more consumption of our core EDA tools. And I think you're starting to see that reflected in our EDA growth this year.

Operator

At this time, there are no further questions. I will now turn it back to Anirudh Devgan for closing remarks.

Anirudh Devgan

Thank you all for joining us this afternoon. It's an exciting time for Cadence as we finish our Q3 with AI-driven product leadership and strong business momentum. The Wall Street Journal ranked Cadence's Board #2 among all S&P 500 companies in its 2026 Best Board list. And on behalf of our employees and of our Board of Directors, we thank our customers, partners and investors for their continued trust and confidence in Cadence.

Operator

Thank you for participating in today's Cadence Third Quarter 2026 Earnings Conference Call. This concludes today's call. You may now disconnect.

原文链接