For years, the hardest job in the chip industry has been invisible from the outside: taking a circuit described in a code-like language and deciding how to lay out billions of transistors on a small square of silicon. Synopsys and OpenAI now say they have a model for that work, one built to operate the software tools themselves rather than simply advise the engineers who use them.
The two companies announced the partnership on September 30, calling the model GPT-Synopsys. It will be trained and tuned specifically for chip design, learning to run Synopsys’s electronic design automation, or EDA, tools the way an expert engineer would, interpreting their output and iterating toward better designs. OpenAI is licensing the tools to develop the model, and the companies will share revenue and go to market together under a multi-year agreement.
The pitch from both chief executives is that the industry’s current bottleneck is not ideas but hours. Synopsys president and chief executive Sassine Ghazi said the goal is to shave weeks or months off the design process and bring more chips to market without giving up on power, performance and area. OpenAI president Greg Brockman framed the project the other way around: by helping engineers build better chips faster, the company can build better AI and reach more people.
The mechanics are what distinguish GPT-Synopsys from the wave of AI assistants already attached to design software. Today, general-purpose models can be connected to EDA tools to run workflows, but the model and the tool remain separate. The new effort is meant to make a frontier model a native user of Synopsys’s tools, so an engineer can hand over a design objective and let agents run the tools, read the results and make changes until the outcome passes review.
Engineers would delegate tasks like power, performance and area optimization, timing closure and verification closure. The model is designed to run on OpenAI-hosted infrastructure and to fit into Synopsys’s existing software, including the Synopsys.ai platform and the Autopilot agentic system. The companies said customer design data will not be used to train the model, will be encrypted in transit and at rest, and can be governed through retention, audit and permission controls.
The announcement did not include pricing, a release date or benchmarks from production work. That gap has drawn attention, because Synopsys has been careful elsewhere to publish measured results. In July it introduced autonomous design workflows built with Microsoft and used by AMD, and said early evaluations of a fully autonomous debug-closure workflow cut cycle time by 25% to 40%. No equivalent number has been offered for GPT-Synopsys.
OpenAI has its own internal data point, though it comes from a different direction. The company said in an August performance report that its Jalapeño inference chip moved from initial design to tape-out in nine months, shortened by faster design, measurement and verification loops. That result speaks to what AI can do inside OpenAI’s own walls; it does not yet show GPT-Synopsys running Synopsys tools inside a customer’s environment.
For chipmakers, the appeal is straightforward. Design teams are stretched, and the number of designs in flight keeps rising as AI drives demand for specialized silicon. A tool that can explore more design alternatives and close on timing and verification faster could let a small team act like a much larger one. The risk is that the model’s judgment still has to be checked by the same experts it is meant to replace, which is why every output is described as flowing through engineer review.
For Synopsys, the deal is a bet that the value in EDA is migrating from the software license to the intelligence that operates it. The company sells the tools that most advanced chips are designed on, and its chief rival, Cadence, has moved into AI-assisted design as well. Tying its tools to OpenAI’s models gives Synopsys a way to sell bundled compute, model access and licenses as a single service, an offering that looks different from a traditional seat license.
The wider contest is over who owns the design loop itself. Chip design has been one of the last parts of the technology stack where expert judgment resists automation, because the trade-offs are unforgiving and the cost of a mistake is a wafer run that produces nothing usable. A model that can navigate those trade-offs would change the economics of an entire industry, and both companies are now spending to be the ones that get there first.
For OpenAI, the move extends its models into a business that touches every other AI ambition. The chips that run AI systems are themselves designed with software, and OpenAI is now selling into that upstream layer while building its own silicon downstream. It is a way to be present on both sides of the same constraint, and it turns the company’s frontier models into a tool that the semiconductor industry, not just consumers, pays for.
