AI is accelerating the EDA tool industry revolution from shallow waters and conceptual stages into deep waters. Discussions around "EDA + AI" used to focus on point-tool optimization or auxiliary enhancements. Today, the continuous evolution of large language models (LLMs) and the maturation of agentic technologies are driving this transformation toward a fundamental reconstruction of design methodologies. EDA vendors are accelerating their layouts, moving from embedding AI capabilities into tools to building agentic systems capable of autonomous decision-making and collaborative execution. Efficiency improvement and design paradigm innovation have become the focal points of competition. Siemens EDA stands out remarkably in this race—according to its latest financial report, its EDA business grew by over 30%, becoming the strongest driver of the group's software business (which grew by 15% overall).
It is precisely at this industry tipping point that the 2026 Siemens EDA Forum was recently held in Shanghai. This annual gathering of supply chain enterprises, technical experts, and partners coincides with a subtle moment in the semiconductor industry: chip design complexity continues to rise, AI accelerators have become standard in chips, and heterogeneous integration is pushing the number of transistors in a single package toward the trillion level. During media exchanges inside and outside the venue, a deeper proposition was repeatedly questioned: as AI evolves from an "auxiliary tool" into an "agent" capable of autonomous decision-making, how will the EDA industry, which has always taken precision and reliability as its bottom line, embrace this transformation?
The answer given by Siemens EDA is—AI-native design.
From AI-Assisted to AI-Native: A Fundamental Innovation in Design Methodology
"Traditional EDA software can no longer meet the demands; future chip design tools must be driven by AI as the core." As early as the opening speech at SEMICON China 2026, Ankur Gupta, Executive Vice President of the IC Products Division at Siemens EDA, made this clear. Eight months later at the Shanghai forum, this judgment was further systematized into the strategic framework of "AI-native design."
Ankur Gupta, Executive Vice President of the IC Products Division at Siemens EDA
What is AI-native? At the communication meeting, Ankur Gupta gave a clear definition to the media, including EEFocus: early traditional algorithms combined with machine learning belong to the "AI-assisted" category; whereas AI-native possesses stronger capabilities in modeling and probability, offering not only faster speeds and broader coverage in design space exploration but also higher accuracy. "Self-verification capability" is the qualitative change brought by AI-native—agents can autonomously verify results during design space exploration, which is unattainable by traditional methods.
Behind this transformation is the continuous evolution of large language models (LLMs). "There is a major version update almost every four months," Ankur Gupta pointed out. It is these advancements that bring continuous momentum to the intelligence of EDA tools.
However, AI-native is not achieved overnight. Pete Ling, Global Vice President of Siemens EDA and General Manager of China, frankly admitted when reviewing this evolution that a few years ago, he repeatedly reminded the industry not to be misled by the hype of "EDA + AI"—"at that stage, the technology was not mature, and AI could not be abused." The chip and semiconductor industry is highly intensive in talent and capital, requiring massive investments and demanding trusted results. "In the past two years, AI capabilities, modeling levels, and self-verification capabilities have reached higher levels, which in turn has enhanced the capabilities of both software and hardware."
Pete Ling, Global Vice President of Siemens EDA and General Manager of China
Three Pillars and Five-Layer Architecture: The Engineering Foundation of AI-Native Design
Under the strategic framework of AI-native design, Siemens EDA proposes "Three Pillars":
- Faster Engines — "Accelerate Every Tool, Every Stage, Every Time." By leveraging AI agents, GPU computing, and physics-based AI solvers, every tool, every stage, and every run is accelerated to the limit, with some processes achieving up to a 1000x speedup.
- Smarter Execution — "More Time for Innovation, Less Execution Overhead." Through agentic AI workflows, intelligent pattern recognition, and context-aware automation, engineers are freed from tedious execution tasks, and overall productivity can be improved by 10 to 50 times.
- Trusted Outcomes — "Release-to-Manufacturing with Confidence, Not Hope." Through lifecycle monitoring, silicon-proven AI models, and comprehensive digital twin technologies, every delivery is ensured with the certainty of first-time-right tape-out, minimizing design risks to the greatest extent.
These three pillars are not mere slogans but have clear technical implementation paths.
From a technical architecture perspective, Siemens EDA has built five technical layers, from bottom to top:
1. Accelerated Computing Layer — Based on hardware such as Arm CPUs or GPUs, utilizing NVIDIA accelerated computing and CUDA-X libraries to "achieve sign-off quality results in hours, not days."
2. Advanced Reasoning Layer — Powered by the NVIDIA Nemotron model to provide advanced reasoning capabilities, supporting scalable agentic workflows and reducing operational costs with higher token efficiency.
3. Secure Orchestration Layer — Leveraging NVIDIA OpenShell for enterprise-grade workflow orchestration, providing security protection, audit trails, and compliance control to ensure autonomous agents operate in a controlled environment.
4. Agent Layer — Dedicated EDA AI agents trained based on NVIDIA NeMo Gym, featuring self-verification capabilities, continuously learning and optimizing to improve task quality and execution speed.
5. Proprietary EDA Engine Layer — Siemens' physics-based simulation engines accumulated over decades, embodying the deep expertise of PhD teams in mathematics and computer science algorithms, providing deterministic, physically accurate sign-off-level results for the top layer.
The core hub of this architecture is the Fuse EDA AI system, released at the NVIDIA GTC in March 2026. Fuse is not just an AI feature of a single EDA tool, but a data, knowledge, and agent orchestration platform sitting above various tools. It builds an ontology for the data generated during product runs through structured foundation models, and on this basis, achieves the orchestration and scheduling of agents.
On July 26, 2026, during the DAC 2026 conference, Siemens and NVIDIA deepened their strategic cooperation by adding self-verification capabilities to the Fuse EDA AI Agent. This means that AI agents performing different tasks can communicate with each other, automatically verify results, and autonomously complete complex processes that originally took a long time. This upgrade is fully supported by the NVIDIA technology stack: NeMo Gym for agent training optimization, OpenShell for secure workflow orchestration, Nemotron models for advanced reasoning capabilities, and CUDA-X accelerated libraries to significantly boost computing efficiency.
Strategic Acquisitions Accelerate AI-Native Layout
While advancing AI-native design, Siemens EDA is also continuously strengthening its capabilities through strategic acquisitions. In July 2026, Siemens announced the acquisition of two companies: Precision Innovations and Defacto Technologies. Precision Innovations focuses on AI-driven design exploration and optimization, significantly shifting feasibility analysis to the left and accelerating digital implementation from RTL to GDS; Defacto Technologies is dedicated to automating SoC (System on Chip) design creation and integration, covering the entire process automation from early RTL to implementation evaluation, verification, and sign-off. These two acquisitions further consolidate Siemens EDA's full-flow technical foundation in the AI-native era.
From Solido to Questa One: Two Facets of Agent Implementation
Beyond the theoretical framework, specific product implementations perhaps better illustrate the point.
In the custom IC (Integrated Circuit) domain, the agent upgrade of the Solido product line provides an intuitive case. Engineers only need to input natural language prompts at the top, such as "characterize all cells," and the system can automatically process it without human intervention—even if the library contains thousands of cells covering multiple process, voltage, and temperature corners. The system will autonomously run simulations, debug results, fix errors, and complete all complex tasks.
More convincing is the striking contrast in efficiency data:
Comprehensively calculated, the characterization workflow is accelerated by over 10 times, and token costs are reduced by 5 to 10 times. Engineers only need to set the intent, turn off their computers, and the next day they can obtain the fully audited, sign-off-ready library files.
In the digital verification domain, the Questa One Agentic Toolkit released on February 27, 2026, is another important milestone. The verification phase typically accounts for up to 70% of the chip design cycle. Questa One automatically generates verification plans and test cases through agents and executes them automatically, transforming verification and design from isolated tool interactions into intelligent, domain-specific multi-step autonomous workflows. The toolkit includes multiple dedicated agents such as RTL code agents, Lint agents, and CDC agents, each performing its own duties while collaborating. With this toolkit, engineers can get started within hours, and tasks that originally required weeks of training are greatly simplified.
Notably, Siemens EDA emphasizes that these agentic workflows are not meant to replace engineers but to enhance their productivity. As Lincoln Lee, Global Vice President of Siemens EDA and General Manager of Technology for the Asia-Pacific Region, stated: "With agents, the focus of engineer training will undergo a fundamental shift. In the past, engineers had to master the specific commands and operational details of EDA tools proficiently. Now, the core responsibility shifts to defining 'what to do' and making professional judgments, while the specific execution process is handled by agents."
Lincoln Lee, Global Vice President of Siemens EDA and General Manager of Technology for the Asia-Pacific Region
Open Ecosystem and Insurmountable Boundaries
In the process of advancing AI-native design, a key proposition is the openness of the ecosystem. Fuse is designed as an open system, supporting the integration of third-party tools and compatible with standards such as MCP (Model Context Protocol). Customers can freely mix and match Siemens' technologies with those of other vendors, and can also develop custom agent skills based on their own expertise.
This openness is particularly critical in the Chinese market. Pete Ling revealed at the communication meeting that since some customers cannot use foreign large models and can only adopt domestic LLMs, they are very concerned about whether the system can integrate local models. "Our answer is yes: the Fuse system is open and supports integrating local models through MCP extensions to achieve customized applications."
However, openness does not mean data sharing. On the sensitive issue of data privacy, Siemens EDA has drawn clear boundaries. Ankur Gupta emphasized that Fuse is deployed in a local environment, and the EDA software runs entirely within the customer's own environment; Siemens does not store customer data. "When customers use third-party tools, the generated log files or final results are also saved locally at the customer's site and are not sent back to Siemens."
A deeper consideration lies in the source of model training data. Lincoln Lee pointed out that although EDA companies do have access to a large amount of customer design data, "we will absolutely not do this, and we must never do this. The designs provided by customers are not for us to use for training; they belong to the customers' core confidential information." Currently, the data used for training by Siemens EDA is strictly limited to its own designs, public datasets, and cases explicitly authorized by customers to be open.
The system also provides a "Bring Your Own Model" (BYOM) feature, allowing customers to perform data training and fine-tuning in their own environments. More importantly, the capabilities accumulated for specific customers have strict isolation—experiences gained from Customer A will not be transferred to Customer B.
The Future of Engineers: Replaced or Liberated?
As AI agents are capable of autonomously completing more and more work, an unavoidable question surfaces: will the semiconductor industry usher in a wave of layoffs?
Ankur Gupta holds the opposite view. He provided several sets of data: in the past, customers used 20 engineers to design a product in 18 months; now, the delivery cycle is required to be halved to 9 months, and the same 20 engineers even need to deliver two products within 9 months. According to forecasts by industry analysis firms such as Deloitte, the total number of engineers needed in the global semiconductor industry by 2030 will be about 1 million. The EDA industry is expected to have a shortfall of 6,000 to 7,000 people by 2030. "Even if AI agents improve productivity, it mainly makes mid-level engineers more efficient and helps junior engineers grow into mid-level engineers faster. But from the perspective of the EDA industry and our customers' semiconductor industry, the overall personnel demand is still rising."
Lincoln Lee responded to this question from another perspective: the semiconductor industry is expected to reach a scale of $1 trillion by 2030, and it is now estimated that it could reach this by the end of this year or next year, and may increase to $1.5 trillion by 2030. "The industry scale is rushing from $1 trillion to $1.5 trillion three years ahead of schedule. Do we have enough engineers? Everyone is excited that the $1 trillion goal will be achieved three years early, but note that we cannot make engineers graduate three years early."
In Pete Ling's view, engineers' skills are transforming rather than being replaced. "In the past, engineers needed to know how to do everything, solving difficult problems and simple tasks personally; now, as long as they are good at asking questions and can clearly describe requirements, a large amount of basic work can be done by AI." The truly difficult work—professional judgment, modeling, and physical-level understanding—still cannot be separated from deep human involvement.
Conclusion: Trustworthiness Remains the Bottom Line of EDA
Looking back at this communication meeting, the word "trusted" was repeatedly mentioned. Whether it is "Trusted Outcomes" as one of the three pillars emphasized by Ankur Gupta, the pursuit of "continuously verifying decisions" by self-verifying agents, or the "engineering implementation" and "yield assurance" repeatedly emphasized by Pete Ling—in the EDA industry, where tape-out costs often involve hundreds of millions of dollars, any technological innovation must take reliability as its bottom line.
Siemens EDA's AI-native design strategy is essentially answering an industry-level challenge: how to compress the chip design cycle from 18 months to 9 months without sacrificing reliability? The path it provides is to make AI not just an auxiliary tool, but an agent capable of autonomous decision-making and self-verification, embedded throughout the entire lifecycle from chip to system. With over 30% business growth, continuous strategic acquisitions, and the large-scale implementation of self-verifying agents, Siemens EDA is marching forward rapidly on this path.
Whether this path can be successfully navigated remains to be tested by time. For the entire EDA industry, however, the curtain of the transformation from the "tool era" to the "agent era" has just been raised.