Cadence embeds AI 'super agent' to assist engineers when designing EDA tools — company aims for ‘over a trillion transistors’ by 2030, with AI helping to debug and verify complex projects

“A trillion transistors” sounds like hyperbole, until it’s not. Designing today’s leading-edge transistors resembles expanding a structure that never stops growing. Chiplets are bolted on as packaging stacks everything vertically, and elements like power delivery and interconnects get re-routed each time footprints change. Eventually, the challenge of designing the actual blueprint is taken over by coordinating everything that connects to it.

That coordination is what Cadence is targeting with what it calls its new ChipStack AI Super Agent. Announced February 10, the system embeds an AI-driven assistant across Cadence’s electronic and design automation portfolio, with the stated aim of helping engineers design, debug, verify, and sign off complex semiconductor projects more efficiently.

“We’re easily going to get over a trillion transistors… in the package, by the end of the decade,” Cadence Senior Vice President Paul Cunningham said in remarks to Bloomberg. “It’s a phenomenal increase in complexity.”

Latest Videos From

Watch full video here:

Emphasis on packaging

Cunningham’s emphasis on “in the package” is significant because growth to a trillion transistors is not simply about shrinking features at advanced nodes, but assembling multiple dies into unified systems through chiplets, 2.3D interposers, and 3D stacking. And as architectural ambition shifts upward, so too does the burden on design orchestration.

EDA has long since automated the mechanics of design, with synthesis, placement, routing, timing analysis, and verification all deeply algorithmic processes that run on large compute clusters. Multi-die packages introduce additional interconnect domains and power islands, push timing closure across die-to-die links, and require thermal modeling across stacked silicon. Verification is therefore no longer isolated to block correctness but extends to system-level interaction between heterogeneous components.

Engineers manage these interactions through layered tool flows, scripts, constraints, and sign-off checks, but reports can easily stretch into thousands of lines, and debug cycles can take weeks. Cadence’s Super Agent sits above that orchestration layer, acting as what the company describes as a domain-trained AI system built on a “mental model” of chip design rather than a general-purpose large language model. Engineers can interact with the tool stack conversationally, requesting actions without navigating every script or endless menus.

“You can chat with all of the Cadence products, and they’ll talk back to you,” he said. “You don’t need to be the ultimate scripting expert. You don’t need to know all of the fancy features and tool clicks of our graphic user interfaces. You can just say, ‘Hey, look. This is what I want to do.’”

Labor constraints

The Semiconductor Industry Authority (SIA) projects that the U.S. could face a shortfall of tens of thousands of industry workers by the end of the decade. A meaningful portion of that gap is expected to fall on engineers and technicians with advanced degrees. While expansion in fab capacity under the CHIPS Act — and corresponding worker shortages — is what tends to draw the most attention in terms of the semiconductor skills shortage, design expertise is equally constrained.

Advanced-node chip design and development demands specialists in physical implementation, verification methodologies, packaging, signal integrity, and system architecture, among other skills. These are roles that require years of specialist training and practical experience, but colleges and universities aren’t currently producing graduates at a rate sufficient to match projected demand and expansion.

So, while it’s all well and good saying “we’re going to break the trillion transistors mark,” the demands on process coordination increase accordingly as more blocks and interfaces translate into more verification scenarios and failure modes. Without significant changes in workflow, the same design processes will demand more and more engineering hours that will be difficult, if not impossible, to fill.

Cadence claims the Super Agent can deliver productivity improvements of up to ten times in certain tasks, a figure that likely reflects targeted gains in repetitive, report-heavy processes. Even so, incremental reductions in iteration time can easily compound across long design schedules and go a long way towards accelerating development cycles.

Agentic chip design

Generative AI has also brought uncertainty to many software categories, raising questions around whether generalized models could abstract away from specialized tools. EDA vendors operate in a tightly constrained market governed by the likes of foundry process design kits and sign-off criteria. Here, things like timing analyzers and physical verification tools aren’t easily displaced by general-purpose models.

But vulnerabilities do exist: For decades, using EDA tools has required fluency in scripting, flow configuration, and the idiosyncrasies of specific vendor ecosystems. That friction has also been part of what keeps customers embedded. If an external AI assistant could sit on top of those tools and translate intent into tool commands, the interface layer would begin to loosen.

By embedding a domain-trained model inside its own stack, Cadence ensures that the conversational interface is part of the platform rather than one of said overlays. This ‘Super Agent’ doesn’t replace timing engines or verification tools, but rather, it sits between the engineer and those tools, interpreting intent and mapping it onto validated flows within the Cadence ecosystem. As architectures fragment into chiplets and stacked dies, constraints are only going to grow, and an assistant that can interpret a designer’s request and translate it into correctly sequenced tool operations might reduce misconfiguration and shorten debug cycles.

Cadence isn’t alone here. Synopsys and others are known to be pursuing similar AI-assisted layers across their portfolios, folding them into tool stacks as a way to reduce coordination overhead. If Cadence succeeds in doing this in a meaningful way, it changes how teams cope with the kind of scale Cunningham is describing.

A trillion transistors in a single package multiplies interfaces, verification, and the number of ways things can fail. More dies equals more cross-domain timing checks. More silicon stacks equal more thermal interactions to model. The coordination workload, as we mentioned earlier, rises alongside that. If the supply of advanced engineers doesn’t expand at the same pace, that coordination burden has to be absorbed elsewhere. Either development cycles stretch, or the effective output per engineer increases. Cadence is trying to make the latter happen from inside the toolchain.