Cadence Bet Against the Hosted Model, Its New Chip-Design AI Runs Behind Your Firewall, and That Is the Point
The two companies that design the world’s chips just split on the biggest question in AI
Almost every advanced chip on earth is designed with software from one of two companies: Cadence Design Systems and Synopsys. Their tools, electronic design automation, or EDA, arrange billions of transistors to meet cost and performance targets, and both firms are now racing to put AI agents on top of those tools so the agents, not just human engineers, do the designing.
On Monday, October 26, Cadence launched its answer: Maestro, a portfolio of AI models tuned specifically for chip design. Reuters reported the launch the same morning. Maestro was built in part on Nvidia’s open-weight Nemotron model and is secured with Nvidia’s OpenShell safety technology. The detail that matters most is not the model. It is where the model is allowed to run: on Cadence’s own cloud, on any public cloud, or, in Cadence’s words, entirely in-house, behind a customer’s own firewalls and inside its own data centres. Customers can also fine-tune the models on their own chip designs.
Four weeks earlier, Synopsys made the opposite choice. On September 30 it announced GPT-Synopsys, a model developed with OpenAI under a revenue-sharing deal, which will run on OpenAI-hosted infrastructure. Same industry, same problem, two opposing bets on where enterprise AI should live. That split, not any single benchmark, is the story.
What Cadence actually shipped
Maestro is a family of models fine-tuned from Nvidia’s Nemotron line to work with, and interpret the output of, Cadence’s own design tools. It sits underneath a set of “AI Super Agents” Cadence has rolled out over the past year, ChipStack for system-on-chip design and verification, ViraStack for analogue circuit design, InnoStack for PHY routing, and AuraStack for advanced packaging and board layout. Those agents orchestrate Cadence’s existing EDA, test and verification tools the way a coding assistant orchestrates a compiler and a test suite: a design either passes verification or it does not, which makes chip design unusually well suited to agents that can check their own work.
Cadence’s senior vice president Paul Cunningham put the current state of adoption plainly: somewhere between 20 and 30 percent of chips under development today are already being designed with the help of AI agents, and the share is growing rapidly. The problem Maestro is meant to fix is that the first generation of those agents ran on off-the-shelf general-purpose models, which had seen little of the material chip design actually runs on, register-transfer-level code and tool output, for example. As Cunningham put it, the powerful out-of-the-box foundation models “are the starting point, and you can still specialize.”
The specialization claim comes with numbers, and they should be read with the qualifier Cadence itself attaches to them. In early testing, Maestro models showed up to 30 percent higher output quality and up to five times better cost-efficiency, The Next Platform, reporting the same launch, phrases the second figure as up to 5x greater token efficiency, versus general-purpose foundation models. The first Maestro models reach select early-access customers in the fourth quarter, so independent results do not exist yet.
One nuance: fine-tuning on a customer’s own designs is real, but Cadence expects to offer it largely as a managed service. “Runs in-house” and “tuned by you alone” are not the same promise.
The Synopsys contrast, in its own terms
GPT-Synopsys is not a weak alternative; it is a different theory of value. Under the September 30 deal, OpenAI pays Synopsys a training subscription fee to learn its tools, the model is trained and tuned for chip-design tasks from circuit description through transistor layout, and when customers use the product, Synopsys and OpenAI share revenue based on how much the model improves the design. OpenAI co-founder Greg Brockman pitched it as a way “to shave off weeks, months from the design process and to bring more chips to the world.” Synopsys CEO Sassine Ghazi said the structure “will not be cannibalizing our business,” and the market liked it: Synopsys raised its fiscal 2027 revenue-growth expectation to 15 percent, above analyst estimates, and its shares rose as much as 7 percent on the announcement. Synopsys also noted the model’s work will still be double-checked by its traditional verification tools.
So the hosted route offers the strongest general model, a partner who profits only when designs improve, and zero infrastructure for the customer to run. The in-house route offers control. Which one wins depends on what a chip design actually is to the company that owns it: not a document, not a dataset, but the crown-jewel intellectual property of the entire business, often the single most valuable secret a semiconductor firm holds, in an industry currently defined by export controls and state-level competition over exactly that knowledge. Sending design work, or even design-adjacent queries, to a model hosted by a third party is a decision many chip firms have been unwilling to make at any capability level. Cadence is betting that for this customer base, deployability beats raw model power.
Cunningham, notably, refused to turn that bet into ideology. He framed Maestro as model-agnostic, “there are more model providers out there than there are chip design software vendors”, and told Reuters that open- and closed-weight approaches are being “pursued equally,” with commercial-lab partnerships “part of the Maestro vision.” In other words, Cadence will happily plug in a proprietary model later. What it will not do is make a hosted proprietary model the only door in.
Why this fork matters well beyond chip design
Chip design is an extreme case, but the same fork is appearing everywhere AI meets genuinely sensitive work, and this month’s coverage on this site has hit it from three directions: dedicated local-AI hardware meant to move agents off the cloud, open-source inference stacks cutting the cost of self-hosting a frontier-class model by an order of magnitude, and IBM telling investors this week that open-weight models now handle routine enterprise work at a fraction of hosted cost. Cadence’s launch is the enterprise-software version of that pattern, and it adds a lesson the hardware stories cannot: the moat was never the base model. Nemotron is open-weight and available to anyone, including, in principle, Synopsys. Cadence’s advantage, if it has one, is the fine-tuning on its own tools’ behaviour, the agent harnesses around them, and a deployment story that lets a paranoid customer say yes. In vertical AI, distribution and permission are becoming worth more than the model itself.
There is also a quieter signal in the partner: Nvidia now supplies the open model and the safety runtime for in-house EDA agents while its chips remain the thing being designed. Every route to designing the next generation of hardware runs through Nvidia’s stack.
The honest caveats
- Every performance number here is a vendor’s “up to.” Thirty percent better quality and 5x efficiency are early, self-reported, and measured against unnamed general-purpose baselines. Early access starts in the fourth quarter; judge Maestro on customer results in 2027, not on launch-day claims.
- In-house still means infrastructure. Running and fine-tuning design models behind your firewall requires GPU capacity, operations staff and patience that the hosted route deliberately eliminates. For smaller design houses, GPT-Synopsys may be the only practical option, which is precisely the segmentation Cadence’s model-agnostic hedge is built to survive.
- Neither approach removes verification. Both vendors are explicit that traditional EDA verification still checks the AI’s work. The agent proposes; sign-off tools dispose.
What to do with this, this week
- If your AI work touches crown-jewel IP, make deployment location a first-class requirement. Ask every vendor, in writing: where does the model run, what leaves our network, and can we fine-tune without our data training someone else’s product? Cadence just proved a major vendor will answer all three in your favour if the market demands it.
- Price the general-purpose tax. A specialised model claiming 5x token efficiency is a reminder that running frontier general models on narrow, repetitive work is often paying for capability you never use. Audit one workflow for a smaller, tuned alternative.
- Keep the model layer swappable. The most durable sentence in today’s launch is Cunningham’s: there are more model providers than EDA vendors. Architect so the model underneath your agents can change without rebuilding the workflow around it, because by next year, it probably will.
The bottom line
Cadence and Synopsys sell near-identical tools to the same customers, and in four weeks they chose opposite answers to the defining question of enterprise AI: does the model come to the data, or does the data go to the model? Synopsys bet on the strongest hosted partner and a share of the upside. Cadence bet that in an industry built on secrets, the winning feature is a firewall. Both may be right for different customers, but only one bet generalises to every bank, hospital and law firm now deciding how much of itself to send to someone else’s cloud.