Cadence Unveils Maestro Semiconductor Design AI

Cadence Design Systems unveiled Cadence Maestro, a family of models for AI agents that perform semiconductor design tasks, on the 26th local time. The company said Maestro delivers up to 30% higher output quality and up to 5x better token efficiency than general-purpose foundation models. Cadence did not disclose which power, performance or area metric underlies the quality figure, or identify the general-purpose models used for comparison.

Some Maestro models were developed from NVIDIA’s open-weight Nemotron models. Cadence said it uses NVIDIA’s OpenShell AI security technology to protect runtime environments on-premises and in private clouds. Customers can run the models through Cadence Cloud or public clouds, or operate them inside their own data centers and firewalls.

Maestro connects with Cadence’s AI Super Agent to interpret design intent, divide work into tasks and coordinate multiple AI models. The system evaluates trade-offs across design requirements and supports decisions. Cadence also said it uses reinforcement learning: deterministic solvers assess model outputs against engineering goals, constraints and correct-answer data, and feed the results into model improvement. Customers can use fine-tuning tools to train models further on their own design data and workflows, using GPUs and networks they manage.

Cadence Senior Vice President Paul Cunningham said costs for AI agents used in design are rising quickly, and that incorporating repetitive design work into model weights is a natural next step. He attributed the claimed improvement of up to 5x in token efficiency to allocating design tools and LLM calls according to each task, instead of using the largest and most expensive model for every job. Cadence did not say whether lower token use would reduce customer bills. Its subscription model includes a set amount of usage and charges extra when customers exceed the limit.

Rob North, senior group director for strategy and new ventures at Cadence, said some design tools take hours to days to produce results. Cadence plans to offer initial Maestro models early to select customers in Q4 2026, but has not announced a general release date.

The announcement differs from Synopsys’ approach. On September 30, Synopsys announced that it would develop GPT-Synopsys, a model specialized for semiconductor design, with OpenAI. That model will run on OpenAI infrastructure, and the companies said customer data will not be used for training. Cadence emphasized that customers can run models in their own environments and use internal data for further training. Both companies cite intellectual property protection and security, but their approaches differ in where design data is stored and how it is used.

Cadence Design Systems unveiled Cadence Maestro, a family of models for AI agents that perform semiconductor design tasks, on the 26th local time. The company said Maestro delivers up to 30% higher output quality and up to 5x better token efficiency than general-purpose foundation models. Cadence did not disclose which power, performance or area metric underlies the quality figure, or identify the general-purpose models used for comparison.

Some Maestro models were developed from NVIDIA’s open-weight Nemotron models. Cadence said it uses NVIDIA’s OpenShell AI security technology to protect runtime environments on-premises and in private clouds. Customers can run the models through Cadence Cloud or public clouds, or operate them inside their own data centers and firewalls.

Maestro connects with Cadence’s AI Super Agent to interpret design intent, divide work into tasks and coordinate multiple AI models. The system evaluates trade-offs across design requirements and supports decisions. Cadence also said it uses reinforcement learning: deterministic solvers assess model outputs against engineering goals, constraints and correct-answer data, and feed the results into model improvement. Customers can use fine-tuning tools to train models further on their own design data and workflows, using GPUs and networks they manage.

Cadence Senior Vice President Paul Cunningham said costs for AI agents used in design are rising quickly, and that incorporating repetitive design work into model weights is a natural next step. He attributed the claimed improvement of up to 5x in token efficiency to allocating design tools and LLM calls according to each task, instead of using the largest and most expensive model for every job. Cadence did not say whether lower token use would reduce customer bills. Its subscription model includes a set amount of usage and charges extra when customers exceed the limit.

Rob North, senior group director for strategy and new ventures at Cadence, said some design tools take hours to days to produce results. Cadence plans to offer initial Maestro models early to select customers in Q4 2026, but has not announced a general release date.

The announcement differs from Synopsys’ approach. On September 30, Synopsys announced that it would develop GPT-Synopsys, a model specialized for semiconductor design, with OpenAI. That model will run on OpenAI infrastructure, and the companies said customer data will not be used for training. Cadence emphasized that customers can run models in their own environments and use internal data for further training. Both companies cite intellectual property protection and security, but their approaches differ in where design data is stored and how it is used.