Synopsys and OpenAI announced a strategic partnership on September 30. The companies signed a multi-year agreement to develop GPT-Synopsys, a specialized model for chip design that brings together OpenAI's frontier AI with Synopsys' EDA tools and chip design expertise. The next leap is to make frontier models experts in using EDA tools: learning to run the tools as expert engineers, interpreting their outputs, and iteratively optimizing designs. OpenAI will pay a training subscription fee, and the companies will share revenue based on how well the model improves customer chip designs. The partnership centers on pairing OpenAI's frontier models with EDA software. The model will train on the tools. The API I call from production systems will not.
The estimating tool at estimate.pro calls OpenAI to extract line items from supplier quotes. The ERP calls Anthropic to parse change orders. The permit tool calls Google to summarize zoning bylaws. None of those models trained on my database schema. None of those models trained on my validation rules. None of those models trained on the PDF parsers I wrote or the webhook handlers I deployed. The models ship with general reasoning capability and I provide the domain context in the prompt. Agentic AI systems currently connect general-purpose models with EDA tools to support chip design workflows; GPT-Synopsys will be developed to use EDA tools, interpret their outputs and iteratively adjust designs. Synopsys is building a model that knows its own software. I am calling models that know nothing about mine.
The specialized model is a product offering not an API endpoint
OpenAI will license Synopsys' chip design software to develop GPT-Synopsys, a specialized AI model the companies plan to sell through a service bundling computing, model access and software licenses. The partnership produces a commercial offering. The offering bundles the model, the tools, and the compute. The customer does not call a public API and paste in tool documentation. The customer does not write a system prompt explaining how to run timing analysis or place transistors. The model is intended to support chip design and verification tasks and operate Synopsys tools directly; engineers will be able to specify objectives such as PPA optimization, timing closure and verification closure, while agents run tools, interpret results, make changes and repeat the process. The engineer specifies the objective. The model runs the tools. The integration shipped with the product.
The construction ERP integrates eight different estimating workflows and four different procurement systems. Those workflows did not exist in any training corpus. Those procurement APIs do not appear in any public documentation the model saw during pretraining. When I call GPT-4 to parse a subcontractor invoice, I send a JSON schema, three examples, and instructions about how to handle sales tax in Massachusetts versus New Hampshire. The model does not know my schema. The model does not know my validation rules. The model does not know that the field called line_total must equal quantity times unit_price unless the override flag is true. I provide that context in every call because the general model has no memory of my tools. Synopsys customers will not provide that context. The context shipped with the specialized model.
The revenue share depends on training access I do not have
The deal is structured so the two companies share revenue when the product is used commercially. Synopsys granted OpenAI access to its software, its documentation, its internal workflows, and its customer use cases. OpenAI will train a frontier model to operate those tools as an expert user. The model will learn how Synopsys engineers run verification. The model will learn how timing closure works. The model will learn the difference between a design rule violation that blocks tapeout and a warning that can ship. That training produces a model that delivers measurable value to chip designers. The revenue share compensates both parties for that value. Synopsys and OpenAI will share revenue based on the improvements the AI model brings to chip design for Synopsys customers. The business model depends on training data I cannot provide and improvement metrics I cannot measure.
The ERP ships invoices to QuickBooks, purchase orders to Procore, and payroll to Gusto. Those integrations move structured data between systems using APIs I built. When I add a new integration, I write the adapter, test the error cases, and deploy the webhook. The model does not learn that integration. The next invoice extraction call knows nothing about the Gusto payroll schema or the Procore purchase order format. I can switch from GPT-4 to Claude or from Claude to Gemini and the integration code stays the same because the model never learned my tools. The API is stateless. The training data is frozen. The specialized model closes that gap for Synopsys customers. The gap remains open for the rest of us.
The model trains on the tools. The revenue share depends on that training. The API call I make tomorrow knows nothing about the database I run today.
Early tests with semiconductor customers are already underway. Synopsys is shipping a product. OpenAI is licensing its models for domain-specific training. The partnership produces a specialized artifact that solves a narrow problem for a specific customer base. That artifact will not appear in the public API. The business-systems operator calling /v1/chat/completions gets the same general-purpose model, the same stateless endpoint, the same context window, and the same requirement to provide all domain knowledge in every prompt. The labs announce partnerships and specialized models while the API surfaces none of that work. The training run sees the tools. The endpoint does not. I keep writing prompts.