AI agents that run chip-design software can burn through a lot of tokens, and Cadence says it has a way to make them use fewer. The company introduced Maestro Monday morning, a portfolio of AI models trained specifically for semiconductor and system design, and said early results show up to 30% better quality of results with up to five times the token efficiency of a general-purpose foundation model.
Maestro sits underneath Cadence’s four “AI Super Agents,” AuraStack, ChipStack, InnoStack and ViraStack, which can route work between Maestro and outside frontier or open-weight models. The models include NVIDIA’s Nemotron, and the two companies are also working on agent security through NVIDIA’s OpenShell reference architecture, which NVIDIA introduced last week.
Where the work happens is the main split with Synopsys. Customers can run the whole setup on Cadence’s own cloud, on public clouds or on premises, and they can post-train the models on their own data, designs and workflows, while Tom’s Hardware reported that Synopsys’s GPT-Synopsys model, built with OpenAI, will run on OpenAI-hosted infrastructure.
Where the savings are supposed to come from
Cadence credits the savings to balancing calls to its design tools against calls to the language model, and its deterministic solvers check the agents’ output against engineering constraints and feed the results back into training. Paul Cunningham, a Cadence senior vice president, told Tom’s Hardware that agent costs have “really started to explode.” He also offered a plainer version of the pitch: “You don’t need to drive to Safeway in the jet engine.”
Tom’s Hardware said Cadence didn’t specify what the quality figure measures, and that the comparison baseline is the unnamed open-source model Maestro started from before its training. Maestro usage also counts toward Cadence’s subscription agent pricing, Cunningham said, and he didn’t say whether the savings would lower a customer’s bill.
Socionext is putting InnoStack into its production design environment, and COO Hisato Yoshida said in the release that the company expects “up to 5X overall productivity gains while maintaining design quality.”
The first Maestro models go to select early-access customers in the fourth quarter, and Cadence hasn’t given a general availability date. Cadence reports third-quarter results after Monday’s close, with its conference call set for 5 p.m. ET.