GPT-Synopsys: OpenAI Is Building an AI Model to Help Design Chips

OpenAI and chip design software maker Synopsys are jointly developing GPT-Synopsys, an AI model meant to operate the industry’s specialized tools like an experienced engineer. The announcement on September 30 targets one of the most expensive bottlenecks in technology: the months that pass between the idea for a chip and the finished design. If it works, the payoff will reach beyond semiconductor companies to everyone waiting for faster, more efficient processors, OpenAI included.

Key takeaways

• GPT-Synopsys is a specialized model designed to run Synopsys design tools on its own, evaluate the results and improve designs step by step.
• Results are still checked by conventional Synopsys tools and ultimately by engineers; the AI does not replace physical verification.
• The model runs on OpenAI infrastructure, and customer data is not supposed to be used for training. Early tests with semiconductor customers are underway, but pricing and a launch date are missing.
• Business model: OpenAI pays a subscription fee for training, and later the two companies share revenue based on how much the model improves a design.
• Rival Cadence is also betting on AI agents in chip design, together with Nvidia; the industry is promising verification cycles of days instead of weeks.

Why chip design is so laborious

A modern processor starts as a description in a code-like hardware language and ends as a blueprint that places billions of transistors. In between lie many steps, all handled with software for what is known as electronic design automation, or EDA. Alongside Cadence and Siemens, Synopsys is one of the largest vendors of these tools. Engineers use them to synthesize circuits, place them on the chip, check signal timing and, above all, verify that the design does what it is supposed to do.

The industry’s goal is called PPA: power, performance and area. Improving one can worsen another, and every run of the tools costs compute time. A lot of engineering time therefore goes into trial and error, interpreting error reports and making further adjustments. That loop is exactly what GPT-Synopsys is meant to take over. According to Synopsys, the model learns to operate the tools the way experienced engineers do. Engineers hand it objectives, from PPA optimization to meeting timing and verification targets; the AI runs the tools, interprets the results, implements changes and iterates until it reaches a verified outcome, which engineers then sign off on.

What OpenAI and Synopsys actually agreed to

The partnership spans multiple years, though neither company has disclosed its length. OpenAI is licensing Synopsys’ EDA tools to train its model; the deal also covers joint research, joint sales and revenue sharing. Synopsys CEO Sassine Ghazi told Reuters that OpenAI will pay a subscription fee for the model to learn how to use the tools. Once customers use the product, revenue will be split according to how much the model improves the chip design. He said the agreement was structured so that it would not cannibalize Synopsys’ own business.

Another of Ghazi’s statements is key to putting the deal in context: the model needs guardrails to check the physics. Every suggestion the AI makes is therefore still double-checked with conventional Synopsys tools, which the industry treats as the authoritative reference for signing off on a design. The AI is a fast operator of the tools, not a substitute for their underlying calculations.

GPT-Synopsys will be offered as a bundle of compute, model and software licenses. It runs on OpenAI infrastructure and is designed to connect to customers’ own agent systems. Because chip designs are among the industry’s most closely guarded secrets, Synopsys stresses that customer data is not used for training and is encrypted both at rest and in transit. Early projects with leading semiconductor customers are already underway, though the company did not name them. Pricing and a date for general availability are also missing. In the announcement video, OpenAI President Greg Brockman spoke of shaving weeks or months off the design process; there are no solid measurements for GPT-Synopsys yet.

A race among the tool makers

GPT-Synopsys does not come out of nowhere. Shortly before, Synopsys had unveiled its Autopilot agent platform and the AgentEngineer solutions built on top of it, with which the new model is to be tightly integrated. They cover six areas, from verification and implementation to manufacturing, and are slated to become available by the end of 2026. According to the company, more than 50 projects are already underway, including with Intel, MediaTek, Samsung and TSMC. Fujitsu reports a 10 to 30 percent productivity gain in generating hardware code, while Synopsys itself cites up to 50 times faster verification. These are vendor figures, not independent measurements.

Rival Cadence is heading in the same direction with its ChipStack AI Super Agent. Around Computex 2026, Cadence extended it with models from Nvidia’s Nemotron family and OpenShell, Nvidia’s security runtime, which we described in more detail when Nvidia introduced its agent watchdogs. Cadence promises more than 40 times faster validation cycles and says a typical five-week verification loop can shrink to less than a day. The difference lies in the model: Cadence relies on Nvidia’s open models, while Synopsys is bringing in a frontier model maker that is training the model specifically on its tools.

For OpenAI, the partnership also serves its own interests. The company is developing its own chips for running AI models together with Broadcom, and demand for computing power is growing rapidly across the industry; the scale of the sums involved is clear from Anthropic’s IPO prospectus, with $518 billion earmarked for compute. Brockman put it this way in the press release: by helping others build better chips, OpenAI can build better AI itself. Investors reacted with confidence: Synopsys shares rose as much as 7 percent after the announcement, which coincided with a revenue forecast above expectations.

Outlook: AI moves to the start of the supply chain

Until now, AI has mostly helped people write software. With GPT-Synopsys and Cadence’s agents, it is moving to the start of the hardware supply chain, where mistakes are especially costly and time is especially short. The business model, with revenue shared according to measurable improvement, is notable: it forces both partners to prove the benefit rather than merely claim it. Whether GPT-Synopsys delivers what Brockman promises will only become clear once the first customers report time savings on real chips. If it succeeds, more and more specialized chips could be built in the future, including by smaller companies that until now have simply lacked the engineers.