Are AI Agents The New Semiconductor IP?

AI & Compute

AI Agents Reshape Chip Design, But a Mistake Still Costs $500 Million

AI agents can deliver customized design solutions in a day versus 10 months for IP, but a $500 million mistake risk at 2nm keeps the industry cautious.

By
Sophie Lindqvist
Filed
Channel
AI & Compute
Read
4 min read

A failed design at 3 or 2 nanometers costs $500 million, and that number explains why AI agents will not replace semiconductor IP anytime soon. "The cost of a mistake is $500 million at 3 or 2 nanometers," said Ankur Gupta, executive vice president of Siemens EDA IC products. "That's why these engines have to remain deterministic. They have to remain high-fidelity."

The remark cuts against the current enthusiasm. AI agents are emerging as a new category of reusable design capability across the EDA industry. Unlike fixed soft IP blocks, agents explore options, integrate results, and refine designs with little human intervention. Some can even compensate for aging hardware. Every promise carries a caveat. An experienced engineering team is still needed to verify what agents produce.

How fast are agents compared to IP?

The speed gap is the strongest claim. "With IP, you need 10 months. Our cycle is much shorter because AI agents are very easy to customize. In some cases we can give solutions to the customer the same day," said William Wang, CEO of ChipAgents. His company runs in-house agents for spec analysis, RTL generation, formal and functional verification, CDC linting, and place-and-route. ChipAgents released a multi-agent solution last year, with parallel agents accessing different contexts for SoC debugging and formal verification.

The parallel is not exact. Soft IP vendors have customized code for large customers for more than a decade to boost performance or cut power. That customization is narrow, costly, and often fails to transfer to derivative chips, especially multi-die assemblies.

Movellus CEO Mo Faisal frames the moment as architectural churn. "I draw parallels to the early 2000s gigahertz wars, where all of a sudden everything was multicore," he said, noting the industry now cycles through architectures every 10 to 15 years.

Who will pay, and how?

Pricing remains unsettled. Cadence has already moved from perpetual licenses to time-based licenses and now leans toward consumption models. Agents "are hitting up the tool usage even more because they want to find answers quickly," said Rob Knoth, senior group director of Strategy & New Ventures at Cadence, making consumption pricing attractive to customers.

Token-based pricing, however, is more talked about than practiced. "People are trying to figure out if anybody in EDA is actually pricing based upon tokens rather than traditional pricing models. It doesn't seem like anyone is," said Wally Rhines, CEO of Silvaco. He sees a model where EDA sells an agent that automates manufacturing-equipment calibration and bills by token count, without either side exposing proprietary data.

What does orchestration demand?

The shift from generative to agentic AI introduces orchestration: a blueprint defining how tasks, tools, and data are coordinated. Coordination fails when multiple agents execute simultaneously with overlapping actions.

Prith Banerjee, senior vice president of innovation at Synopsys, describes the new hierarchy. Agents now cover EDA tools and CAE tools, including simulation with three core solvers — CFD, HFSS for electromagnetics, and mechanical structural analysis. "What was happening by a team of engineers coordinated by a human manager is now being done by a hierarchy of agents," he said.

Keysight's senior vice president Niels Faché calls the implication disaggregation. Platforms are sticky, he concedes, but customers ultimately want flexibility. Keysight is modularizing products through MCP servers, with APIs coded in Python, so agents can stitch together code-based workflows. Front-runner customers already build such workflows, he said.

Arm's Robbie Williamson, vice president of customer engineering for cloud AI, notes the load shifts to CPUs. Hyperscalers now ask what the CPU must deliver for agentic workloads, unlike training jobs that simply batch and wait.

Where does trust break down?

Stelios Diamantidis, chief product officer at Cognichip, warns of ping-pong effects and convergence issues when agents disagree, and of "chaos times chaos" in heterogeneous multi-die stacks built on models with unknown training sources. Gupta adds a data problem: foundries restrict PDK access tightly, and the design universe is siloed, starving LLMs and SLMs of the physics data they need.

Agents will not replace traditional IP. They will change how IP is used, and adoption will be slow in an industry where a $500 million program failure remains, as Faisal put it, a big deal.

Source: Semiconductor Engineering

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,055 articles

Related articles

« Previous articleNext article »