Last year at Cadence LIVE 2025, the company was already heavily engaged in building a variety of solutions that leveraged AI — including agents. At this year’s event, Cadence reiterated its approach to AI in a multi-phase, 10-year plan, starting with infrastructure AI, then physical AI, and finally scientific AI. Cadence CEO Anirudh Devgan kicked things off this year by discussing the semiconductor explosion connected to AI and projections that the industry will reach $2 trillion by 2030. Cadence’s view is that there is an engineering bottleneck caused by the increasing complexity of these new AI chips, and that simply hiring more people isn’t the solution. Cadence says that hyperscalers claim each new iteration of an AI chip requires twice as many engineers, creating an unsustainable trajectory.

Cadence’s solution to this demand issue is to shift from traditional script- and GUI-driven design flows to agentic AI workflows, which it says significantly compress design cycles and improve productivity. NVIDIA CEO Jensen Huang, in his fireside chat with Devgan at the conference, mentioned that, thanks to Cadence’s AI Agents, NVIDIA has already harnessed “infinitely more” chip designers — virtual ones — that will drastically scale productivity. He also mentioned that NVIDIA’s engineers are actively working with Cadence’s “virtual engineers” in Slack channels, a process he says has already helped to convert codebases in five hours that would have traditionally taken 10 engineers an entire year.

Cadence’s Three-Layer Agentic EDA Cake

“Layer cake” metaphors have become popular again in the age of AI, and Cadence has rolled out its own AI platform cake. The base layer is the compute and the infrastructure — the foundational hardware running the operations. The middle layer is Cadence’s core competency in computational software with highly accurate mathematics- and physics-based EDA algorithms. The top layer is all of the AI and Super Agents (more on these proprietary tools below) that deploy reasoning language models and domain-specific agents to properly select the right EDA tool agent to solve the problem as the LLM understands it. Huang reinforced the idea that agentic systems rely heavily on tool use and that AI agents must be grounded in these existing, accurate Cadence tools so that human engineers can easily verify and collaborate with the AI’s outputs.

JedAI, the Invisible Plumbing of Cadence’s Agentic AI

JedAI is Cadence’s underlying orchestration and infrastructure layer, the connective fabric that sits between Cadence’s internal products and external LLMs. According to the company, JedAI is designed to be extremely flexible, enabling customers to use cloud-based models or opt for private, on-premises models to ensure that their proprietary design data remains secure. JedAI handles crucial behind-the-scenes tasks like traffic routing, security, and embeddings so that users can focus on interacting directly with Cadence’s Super Agents.

Cadence also announced at Cadence LIVE that it was expanding support for NVIDIA GPU and Arm architectures, giving customers more choice in the type of infrastructure they run Cadence tools on. Cadence claims that GPU acceleration can deliver anywhere from 5x to 1,000x performance gains compared to running on x86 CPUs.

Cadence’s New Super Agent Portfolio: Automating the Chip Design Lifecycle

The new Cadence AgentStack serves as the head agent in Cadence’s new portfolio of Super Agents. The AgentStack orchestrates multi-agent workflows, allowing certain tasks and tools to be engaged in parallel or switching between Super Agents when necessary, removing silos based on specific agents or toolchains. These new Super Agents build on the ChipStack AI Super Agent that Cadence announced earlier this year, based on technology from its acquisition of ChipStack late last year. ChipStack is specifically focused on front-end RTL generation and verification, and its technology was praised by Huang for understanding complex interfaces like PCIe and CXL, and helping to automate the frustrating process of verification. Cadence said that customers including Altera, Google Cloud, MediaTek, NVIDIA, Qualcomm, and Tenstorrent are already using Chipstack to design their products. Cadence also announced that it was working with Google to optimize ChipStack for use with Gemini on Google Cloud. This gives Cadence not only world-class model support, but also the scale and TPU/GPU infrastructure of Google Cloud.

In its new portfolio, Cadence announced ViraStack, InnoStack, 3DStack, and SystemStack Super Agents. These Super Agents all address different domains of Cadence’s tools: custom and analog, digital implementation and signoff, 3D-IC design, and multiphysics. Huang also praised these tools, specifically saying that Cadence’s analog design operates at the “limits of physics,” which he said makes it an ideal domain for agents to rapidly explore thousands of permutations for signal integrity, power, and size.

Cadence Deepens Its Relationship with NVIDIA

Continuing the NVIDIA theme, Cadence also expanded its partnership with the chip maker by integrating Cadence’s solvers with NVIDIA CUDA-X and AI physics in Cadence’s Millennium M2000 Supercomputer, achieving, it said, as much as a 100x speedup. NVIDIA is also working as an early partner by deploying Cadence’s new AgentStack within its own semiconductor design process.

The integration with NVIDIA doesn’t end there. The companies are also working together for physical AI applications to help close the “sim-to-real” gap for robotics and autonomous machines. Cadence is integrating its physical AI stack with NVIDIA’s Isaac simulation libraries and Cosmos open-world models. This is meant to provide an end-to-end workflow for training, validation, and real-world deployment on NVIDIA Jetson systems. Last but not least, Cadence is also using NVIDIA’s Omniverse DSX Blueprint, announced at GTC, to help create digital twins of massive AI datacenters, specifically those built for the Grace Blackwell and Vera Rubin systems. NVIDIA has integrated Cadence’s Reality platform into its DSX platform, and that’s a big deal for both companies, as the simulations are helping optimize for token burn, supposedly yielding 17% to 32% more tokens-per-watt compared to the last generation.

The Future Trajectory of Silicon Engineering

Cadence is showing the semiconductor industry a path into an era in which human engineers are heavily augmented by an interconnected ecosystem of AI agents. Tools like Claude Code, Codex, and Cursor are proving that the semiconductor industry is not alone in adopting this trend, either. The JedAI platform and Cadence’s suite of Super Agents show that Cadence isn’t only addressing the immediate engineering bottlenecks, but also helping to lay the groundwork for the next phase of AI adoption: life sciences and biological simulation. Even more so than today, this next phase of AI will depend heavily on advances in semiconductor capabilities and rapid improvements in performance and efficiency. That’s what Cadence intends to support.