At Computex 2026, two competing visions for autonomous chip design emerged as Cadence and Synopsys showcased fundamentally different strategies for deploying AI agents in semiconductor engineering. Cadence is betting on depth, pushing verification autonomy to what it calls Level-5 and achieving over 40 times faster RTL (register-transfer level) validation. Synopsys is betting on breadth, extending autonomous workflows beyond code-like front-end tasks into thermal and physical closure, the part of the design flow where costly respins actually originate.

What Does Level-5 Autonomy Actually Mean for Chip Design?

Cadence extended its ChipStack AI Super Agent to Level-5 autonomy at NVIDIA GTC Taipei, running on NVIDIA Nemotron models and the OpenShell runtime. In the workflow highlighted during NVIDIA's keynote, the agent orchestrates Cadence Xcelium simulation and Jasper formal verification, cutting a typical five-week validation loop to under a day. This speed gain matters because verification is the closest chip design task to software development, where AI coding agents already work well. NVIDIA's endorsement carries weight; Jensen Huang mentioned hiring hundreds of thousands of Cadence super agents, giving the depth play a powerful reference customer.

However, hardware engineers should treat autonomy-level claims with caution. Cadence's Level-5 and Synopsys's L4 are self-graded labels on scales that are not standardized across vendors, making direct comparison difficult. Both companies are showcasing pieces of a flow rather than fully integrated end-to-end systems, and named customer references with audited metrics remain thin on both sides.

Why Is Thermal and Physical Closure the Real Bottleneck?

Synopsys countered Cadence's verification focus by demonstrating an autonomous AI engineer that uses Ansys IcePak to mesh, simulate, and optimize GPU cooling inside NVIDIA NemoClaw at Computex. This move extends autonomy into the hardest and least automated part of the design flow. Thermal and physical closure is where respins originate, especially in 3DIC (three-dimensional integrated circuits), making it the domain where automation delivers the most practical value.

"Engineers defining intent while agents orchestrate execution across the full design flow," explained Thomas Andersen, Vice President of AI and Machine Learning at Synopsys, framing the direction for agentic chip design.

Thomas Andersen, Vice President of AI and Machine Learning at Synopsys

Synopsys's Multiphysics Fusion technology folds Ansys thermal, voltage drop, and electromagnetic engines into implementation tools, allowing teams to measure physical effects accurately enough to cut the safety margins they would otherwise build into designs. An agent then runs optimization loops on top of that higher-accuracy analysis. This approach addresses a structural gap: Cadence does not have an Ansys-equivalent multi-physics stack, which explains why the two vendors are pursuing such different roadmaps.

How Are EDA Vendors Building a Moat Against Frontier AI Labs?

Both Cadence and Synopsys describe the same competitive advantage over frontier AI labs. A general coding agent pointed at raw EDA tools can read log files and write skill scripts, but it produces lower-quality, less repeatable results and burns far more tokens. The real moat is tight tool integration, special machine interfaces, and structured data that feed the right context to each model call.

• Tool Integration: EDA vendors have decades of optimized workflows and APIs that frontier labs cannot replicate quickly, allowing agents to work more efficiently with fewer token calls.
• Structured Data: Token-efficient structured data formats give EDA incumbents a durable advantage, since general agents lack the domain-specific context needed for autonomous chip design.
• Grounded Orchestration: The defensible layer is the grounded tool chain and orchestration logic, not the underlying language model, which is why EDA incumbents rather than frontier labs are positioned to own autonomous chip design.

Agent count is becoming a competitive metric, echoing the agents-per-rack metric now governing data center economics. The reasoning model advances of the past year made agents viable, but they did not commoditize the toolmaker.

What Model Improvements Are Accelerating the Timeline?

NVIDIA's Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model, claims roughly 5 times faster inference and about 30 percent lower cost than leading open models. Each step like this lowers the token cost that gates continuous agent use while raising the autonomy a vendor can credibly offer. More design domains, more cross-vendor collaboration, and cheaper, stronger reasoning models turn agentic EDA from a verification accelerator into a consumption-based growth market.

How to Evaluate Agentic EDA Claims as an Industry Participant

• Demand Named References: Ask vendors for named customer references and repeatable, audited metrics rather than vendor-cited multipliers on both verification and multi-physics closure, since marketing claims often outpace real-world results.
• Assess Runtime Bottlenecks: Once agents automate the human analysis loop, the runtime of the underlying EDA tools becomes the limiting factor, so GPU acceleration of simulation, implementation, and signoff is the quiet variable that will decide who actually delivers speed at scale.
• Watch for Standardization: Track whether the industry converges on a shared L1 to L5 autonomy scale or each vendor keeps self-grading, since standardized metrics are essential for comparing competing solutions.

The commercial opportunity scales with two variables. The first is what early adopters do next. Cadence names MediaTek, NVIDIA, Altera, Qualcomm, and Tenstorrent among its first ChipStack users. If those teams push agents beyond front-end verification into analog, 3DIC, and the collaborative, multi-vendor chiplet packages that increasingly define advanced designs, the addressable work for agentic EDA expands far past where it started.

The second variable is model improvement. Nemotron 3 Ultra and similar advances lower the token cost that gates continuous agent use. More design domains, more cross-vendor collaboration, and cheaper, stronger reasoning models turn agentic EDA from a verification accelerator into a consumption-based growth market, and the incumbent that owns the grounded tools captures most of it.