Cadence Q3: The AI Model Isn't Where Chip-Design Money Goes

I think the AI model is the cheapest part of a chip-design agent, and the market priced it as the prize. When Synopsys and OpenAI announced GPT-Synopsys on September 30, Synopsys shares closed up 4.78%. On its October 26 call Cadence described a different setup: small models of its own, run in the cloud or inside the customer's building, with customers free to plug in their own. Whichever model drives the agent, the chip still has to pass through the design software the foundry has certified, and Cadence and Synopsys sell most of it.

Some plain terms first. Chip-design software is called EDA. Before a design goes to a factory like TSMC (the step called tape-out), it has to be laid out and checked on tools the factory has certified for that manufacturing process. An AI agent can do an engineer's job inside those tools, but it can't skip them.

The real difference is the frontier model and where it runs. Synopsys is betting that a model built with OpenAI will do the work best, and it is sharing revenue to get one. Cadence is betting that chip designers would rather keep their designs off someone else's servers, and that small models trained on its own tools are good enough. CEO Anirudh Devgan said Cadence wants "to control the three layers of the cake," meaning the model, the agent and the tools, and that Maestro's training cost "should not affect any financial model of Cadence" because Cadence saves money using AI internally.

So why does Cadence bother building models if they're the cheap layer? I read Maestro as the price of keeping the customer's data on the customer's premises, rather than a moat of its own. It costs Cadence little, and it means a chip team never has to choose between Cadence's tools and its own security rules. Where the two companies agree matters more: both still sell tool licences under the agent, and both expect agents to need more of them.

The quarter was strong in every product group. Core EDA, which is design and verification software plus emulation hardware that had a record quarter, grew 20%. Non-GAAP operating margin fell to 44.1% from 47.6%, which management partly tied to the Hexagon design unit it bought this year. None of this is agent revenue yet. Cadence counts more than 100 agent engagements with customers and discloses no revenue from them, so Q3 is the starting line for the claim, not proof of it.

Recurring revenue is the number to watch because it is mostly software licences, where an agent's work lands. CFO John Wall's argument is that agents "could perform much more design exploration than an individual engineer," and every option explored is another run of Cadence's tools. If he's right, this line keeps climbing as agents spread. If agents find good designs in fewer runs, it bends.

The reason I think the tools keep the value is the certification loop. TSMC, Samsung and Intel certify each EDA vendor's toolchain for every new manufacturing process, designers need a certified toolchain to tape out, and a newcomer can't get certified early enough to matter. An agent from OpenAI, Anthropic or a chip company's own lab still has to run that toolchain. Autodesk and PTC, which sell design software for buildings and machines, are a useful contrast, because part of their moat is people trained on AutoCAD, Revit or Creo, and an agent driving the software through its APIs doesn't need that training.

InvestMoat rates each company's moat pillars as strong, intact, weakened or destroyed, or not applicable. The grid shows the difference in one place.

Cadence and Synopsys have no trained-user pillar to lose, and the foundry network and tape-out embedding are the two pillars the framework rates strong for both. Autodesk and PTC have the trained-user pillar, and the framework still rates it intact. My view is that agents will wear it down over time, which is a judgement about where the framework is heading for those two, not something the grid shows today.

The strongest objection is that I'm asking the wrong question. Nobody serious expects an AI lab to get its own toolchain certified at TSMC soon, so the pair isn't what's at risk. The split between them is. If GPT-Synopsys finishes designs faster, or produces chips that use less power and space than Maestro manages, a chip team picks Synopsys for the agent and buys the tools along with it. In that world the model doesn't break the moat. It decides which of the two certified vendors collects the licence, and teaming up with the best frontier lab was the smart move. The market's reaction on September 30 says plenty of investors think so.

The second objection comes from Cadence's own design. Letting customers bring their own LLM means a chip company can build an agent on a general model and point it at whichever tools it likes, from either vendor, one step at a time. That loosens the full-toolchain bundle, which the framework already rates intact rather than strong because each company sells an equivalent set. An agent that mixes tools freely is a better deal for the customer than for Cadence.

The third is the low end. Moonshot reportedly said its Kimi K3 model designed an inference chip in 48 hours on free, open-source EDA tools. It was an old 45nm process, tested in simulation and never made, but it shows an agent doesn't need certified tools for simple chips, and simple chips are where open-source tools could start taking work.

The fourth goes straight at my thesis: agents might reach good designs in fewer runs, so an engineer's productivity gain becomes a smaller bill. Synopsys's own EDA growth was 8.5% in its July quarter, against a 16% quarter a year earlier, and its long-term plan assumes EDA gets back to mid-teens growth, a target agents now have to help deliver. I can't rule this out. My answer is history: Devgan argues chip design got about 100 times easier over 20 years and the industry ended up doing more designs, not fewer. I read that as more tool use, but that's my inference. That's an argument, not evidence about agents, which is why recurring revenue is the test.

I still hold the view, narrowed by the first objection. The foundry loop protects the two vendors as a pair, and I don't see the model threatening that at the leading edge. Whether Cadence or Synopsys picked the better model strategy is a share question, and one month after GPT-Synopsys is too early to answer it. For a Cadence owner, that means the risk to watch is share, not survival.

• Cadence's recurring revenue growth each quarter of 2027 against the 18% in Q3 2026. The falsifiable claim trips below 10% in two quarters once agents are in production.
• The first agent revenue either company discloses, and whether it shows up as more tool licences or as a separate line.
• Whether GPT-Synopsys moves from early engagements to production at a leading chip designer, and whether any of them say they chose Synopsys because of the agent.
• Cadence's 2027 outlook in February, which Devgan said is when it will update investors.