CITIC SEC: EDA value is expected to be revalued; recommend paying attention to related domestic companies.

CITIC SEC released a research report stating that AI EDA is becoming the next important direction for AI Agent implementation, EDA value is expected to be revalued, and recommends paying attention to domestic EDA companies. EDA is gradually evolving from a "tool" into an "intelligent agent," becoming the Agent operating foundation for the entire chip design process. EDA leaders such as Synopsys have made progress, and AI EDA has begun to enter the validation phase. AI EDA drives value enhancement for EDA companies, with triple upgrades in business model, barriers, and demand. Domestic EDA companies have launched AI efficiency tools, and engineering data has become the core moat. CITIC SEC's main views are as follows: EDA is gradually evolving from a "point tool" into an "intelligent agent." The essence of AI EDA is to use AI agents to take over the most labor-intensive "design-verification-modification" loop in chip design: from early AI-assisted PPA optimization, to generative Copilot helping engineers write code, to today's Agentic EDAwhere multiple AI agents autonomously orchestrate the complete design process. According to Synopsys' official website, Synopsys classifies AI EDA autonomy into L1 to L5 levels based on autonomous driving logic: L1 is knowledge assistance, L2 is action-oriented agents that can autonomously complete specific tasks, L3 is multi-agent collaborative planning, and L5 is fully autonomous; this year, L3-L4 level multi-agent workflows have been demonstrated, such as natural language generation of RTL, automatic verification convergence, etc., with measured efficiency improvements of several times. This means AI is no longer an isolated point tool, but is becoming the Agent operating foundation for the entire chip design process. EDA leaders such as Synopsys have made progress, and AI EDA has begun to enter the validation phase. According to Synopsys' official website: On September 28, Synopsys released seven long-duration agents of AgentEngineer and the Autopilot unified platform. AgentEngineer is a set of domain-specific long-horizon agents capable of "reasoning, planning, and execution" covering the complete engineering process of verification, implementation, simulation and mixed-signal design, manufacturing, etc. Measured data shows verification convergence up to 50 times faster, coverage improved by 20%, productivity improved by 30%, Token efficiency improved by 2 times, and latency significantly reduced. On October 1, Synopsys signed a multi-year binding agreement with OpenAI to jointly develop a chip design-specific model GPT-Synopsys and establish a revenue-sharing mechanism. The model is trained as a "native expert user" of EDA tools, able to directly operate tools, interpret outputs, and iteratively optimize designs. Engineers can delegate design goals such as PPA optimization and timing convergence, and the agent autonomously runs tools and delivers verified results. On October 10, Synopsys officially announced cooperation with Amazon. The two parties signed a multi-year agreement worth more than $1 billion, with cooperation extending from IP licensing to cloud and AI-driven engineering, and for the first time launching a "license fee + royalty" modelas Amazon's self-developed chips such as Graviton, Trainium, and Nitro enter mass production and shipment, Synopsys is expected to receive continuous royalty income. Driving value enhancement for EDA companies, with triple upgrades in business model, barriers, and demand. 1) Model transformation: from selling licenses to a three-tier monetization system of "Agent subscription + computing power consumption + outcome sharing." Taking GPT-Synopsys as an example, OpenAI pays tool license subscription fees for training the model, and after the product is launched, both parties share revenue based on the model's improvement in chip design; 2) Barrier enhancement: large models are general capabilities, but the real moat is the engineering knowledge, process data, signoff flows, and trusted verification engines accumulated by EDA companies over decades. AI agents must rely on deterministic EDA engines for signoff verification. 3) Demand expansion: significant improvement in design efficiency will stimulate more design projects. Global tech giants are collectively developing their own chips, and combined with the leap in Chiplet/3DIC complexity, EDA has entered a new upward cycle. According to Synopsys' financial report, it has raised its long-term revenue growth target to mid-teens, and expects Non-GAAP operating margin to reach approximately 50% in fiscal year 2030. Domestic EDA companies have launched AI efficiency tools, and engineering data has become the core moat. Empyrean Technology: AI matrix-style penetration, Agentic platform positioning. According to company announcements, the company has launched the AI-driven process diagnosis platform Vision, the agent Aether Coder, etc., with defect identification accuracy reaching 99% and efficiency improved by 6 times compared with manual work. The full-custom memory circuit platform Aether achieves simulation back-annotation rate improvement of 20%+, full-layout speedup of 5 times, and critical path tracing speedup of 6 times; the advanced packaging platform Storm engine performance is accelerated by 15-30 times. According to Chairman Liu Weiping's speech at the 2026 IICIE Conference, the company clearly defines 2026 as the year when "Agentic AI becomes an industry consensus" and is building a multi-agent collaborative intelligent agent platform. Primarius Technologies: AI-driven DTCO, enabling full-link automation. The company takes "AI empowering DTCO" as its strategic direction, advocating systematically shortening iteration cycles through full-link collaboration from process, device, model, PDK to circuit simulation. The new-generation SPICE model automated extraction platform MS-SDEPM integrates the full process of model analysis, parameter extraction, and quality verification, efficiently improving modeling accuracy and efficiency through target collaborative optimization. The new AutoCell standard cell layout automation platform realizes full-process automation from layout design to verification, effectively addressing the bottleneck of lengthening design cycles caused by advanced process scaling. Semitronix Corporation: Using a "knowledge base + agent" architecture to transform yield analysis from "people finding answers" to "answers finding people." The SemiMind semiconductor large model platform deeply collaborates with models such as DeepSeek-R1 to realize professional agents for Design Rule checking, automatic test plan generation, and autonomous YMS/DMS analysis. The SemiClaw enterprise-level digital employee platform transforms EDA processes from "people finding information and writing scripts" to "task-driven, agent autonomous orchestration and operation." QuickRoot, with AI one-click root cause analysis, improves defect troubleshooting efficiency by dozens of times, compressing complex analysis processes that originally took weeks into several hours. Risk factors: Company technological innovation, product upgrades, and performance growth falling short of expectations; AI application progress falling short of expectations; downstream demand falling short of expectations; industrial policies falling short of expectations; intensifying industry competition; risk of intensified international geopolitical tensions.