On September 30, OpenAI and Synopsys announced a multiyear strategic partnership to co-develop "GPT-Synopsys," an AI model specialized for semiconductor design. The aim is to let AI operate design software, read its analysis results, and repeatedly revise and verify a design.
Two days earlier, Synopsys had unveiled "Autopilot," an AI platform that supports design work. On September 30, it also announced an IP agreement with Amazon worth more than $1 billion. By combining a dedicated AI model, a platform for automating design, and reusable circuit IP with its existing EDA tools, Synopsys is trying to change not only how chips are designed but also how it charges for its products.
Handing tool operation and revision to AI
GPT-Synopsys is an AI model specialized in using Synopsys's EDA (electronic design automation) tools. EDA is the software that supports chip development from circuit design through functional verification to final pre-manufacturing checks.
According to the companies' announcement, OpenAI will be licensed the EDA tools it needs to develop the model, and the two companies will cooperate on both R&D and sales. The agreement also includes a mechanism for sharing revenue from the joint service.
It is already possible to connect a general-purpose AI model to EDA tools and have it carry out design tasks. This partnership goes further by developing a dedicated model adapted to the way EDA tools themselves are used.
In the envisioned workflow, the AI reads analysis results, decides what to fix next, and improves the design. Details such as which base model will be used and how it will be trained or optimized have not been disclosed.
The goals engineers can give the AI include PPA optimization, timing closure, and verification closure.
PPA refers to three metrics: power consumption (Power), performance (Performance), and area (Area). In design, these must be balanced according to the application and the required specifications.
Timing closure is the work of fixing problem spots so that signals in a circuit arrive within the required time. In verification, too, bugs and gaps in coverage must be reduced until the design satisfies the required conditions.
With GPT-Synopsys, an AI agent runs the EDA tools, analyzes the results, revises the design, and verifies it again. This cycle repeats, and engineers check the final result.
What the two companies announced is the joint development of an AI model that supports this sequence of design work. It is not an announcement that AI alone can now design a finished chip without human involvement and send it straight to manufacturing.
Before a design can move to manufacturing, it needs "sign-off," a final check that it meets the required specifications. Synopsys's existing sign-off product lineup analyzes signal timing, power, and other factors to confirm that a design meets the conditions for manufacturability.
Even if an AI model can generate plausible fixes and explanations, that alone does not prove that a circuit satisfies physical and electrical conditions.
On this division of roles, Synopsys CEO Sassine Ghazi told Reuters in an exclusive interview that design results produced by GPT-Synopsys will be checked again using Synopsys tools that rely on conventional computational methods.
The AI proposes fixes, EDA tools verify the results, and engineers make the final check. Even if AI increases the number of design candidates that can be tried, the step of confirming their correctness does not disappear.
How do Autopilot and GPT-Synopsys differ?
Synopsys has already introduced several AI technologies into chip design.
DSO.ai, available since early 2020, uses reinforcement learning to explore many design candidates and optimize PPA.
GPT-Synopsys, by contrast, is positioned as a model specialized in understanding design and verification goals and iterating on work while using EDA tools. No performance tests comparing the two under identical conditions have been published.
In addition, Autopilot, announced on September 28, is a platform for managing work done by AI agents.
It coordinates multiple processes, uses skills and memory, and manages execution history and permissions. On top of it, "AgentEngineer" is the group of domain-specific agents that handle individual tasks such as verification and implementation.
Customers can choose not only Synopsys models but also models and infrastructure from partners and third parties.
The two companies plan to integrate GPT-Synopsys deeply into Synopsys.ai and Autopilot, and to design it so that it can work with the agent-execution platforms customers already use.
Two days after the September 28 Autopilot announcement, on September 30, Synopsys announced the dedicated-model development with OpenAI and an application-specific IP agreement with Amazon. The three initiatives announced in 2026 can be organized by role as follows.
| Announced | What was announced | Role | Disclosed status and terms |
|---|---|---|---|
| Sept. 28 | Autopilot / AgentEngineer | Agent execution platform and domain-specific design support | More than 50 technical collaborations underway; availability planned for the end of 2026 |
| Sept. 30 | GPT-Synopsys | AI model specialized in using EDA tools | Multiyear joint development agreement; early technical collaboration with customers has begun, but general availability date not disclosed |
| Sept. 30 | IP agreement with Amazon | Supply of circuit design assets optimized for specific uses | Multiyear contract worth more than $1 billion; business model combining licensing and royalties |
The Autopilot platform, the GPT-Synopsys dedicated model, and the circuit IP provided to Amazon each play a different role.
This table organizes where each announcement fits; it is not a comparison of performance under identical conditions. The "end of 2026" availability shown for Autopilot cannot be read as the release timing for GPT-Synopsys, nor can it be linked to any adoption of GPT-Synopsys by Amazon.
Customer results have, however, been published for the existing AgentEngineer.
In the September 28 announcement, Fujitsu said that using Verification AgentEngineer for RTL code generation produced productivity gains of 10–30%. The tasks cited include writing verification conditions in SystemVerilog, generating modules, and tidying code.
This is a result for a specific stage involving RTL, which describes circuit behavior; it does not mean that overall chip development time was cut by 10–30%. Nor can this figure be treated as the performance of GPT-Synopsys.
A $1 billion-plus Amazon deal widens revenue sources
The Amazon agreement further expands the uses of the IP that Synopsys provides.
IP here means reusable circuit design assets that can be built into chip designs. Amazon is developing its own chips, including Nitro, which handles cloud security, networking, and storage processing; Graviton for general-purpose computing; and Trainium for AI training and inference.
With Amazon as a major customer, Synopsys will expand its IP business optimized for specific uses and system configurations.
The contract is worth more than $1 billion and spans multiple years. That is the size of the IP agreement with Amazon; it is neither the value of the OpenAI contract nor Synopsys's annual revenue.
Synopsys has indicated it will adopt a business model that combines IP licensing fees with royalties, so that both companies profit as production volumes rise. Details such as the upfront payment and royalty rate have not been disclosed.
The cooperation between the two companies extends beyond IP supply to design and analysis.
Amazon will expand its use of EDA, simulation, and AI agents. The companies will work together to speed up Synopsys's multiphysics analysis software on Trainium and Graviton, and Synopsys itself will adopt Amazon EC2 and Bedrock in its own product development.
However, the execution environment announced for GPT-Synopsys is one hosted by OpenAI. The Amazon partnership does not support concluding that GPT-Synopsys will be offered on AWS.
The OpenAI partnership also includes a mechanism through which Synopsys earns revenue from use of its EDA tools.
According to Reuters, Ghazi said OpenAI will pay subscription fees for using Synopsys's tools. When customers use the joint service, the two companies will also share revenue according to the design improvements the model delivers.
Specific terms, such as how fees will be calculated and which metrics will be used to evaluate design improvements, have not been reported.
In its Investor Day announcement the same day, Synopsys said it would combine subscription and usage-based pricing across EDA tools, AI agents, and the execution platform.
It envisions multiple usage models: customers who adopt Autopilot as a package, customers who build only Synopsys's agents into their own AI infrastructure, and customers who use a dedicated model such as GPT-Synopsys.
Once AI takes over repetitive work, the time engineers spend working and the number of times EDA tools are run will no longer necessarily be proportional.
Taken together, the announcements show that Synopsys is trying to earn revenue not only from traditional design-software license fees but also from AI-driven design work and wider use of application-specific IP. How much this will actually contribute to profit will depend on how widely the services are adopted and on contract terms yet to be revealed.
Judging practicality means looking at design time, including verification
The GPT-Synopsys joint service is planned to bundle computing resources, the AI model, and EDA tool licenses.
The two companies say customer design data will not be used to train the model and will be encrypted both at rest and in transit. They also say data retention periods, auditing, and access permissions can be configured.
Semiconductor design data is confidential information directly tied to a company's competitiveness.
Beyond the assurance that data won't be used for training, companies will need to confirm whether processing on external computing infrastructure meets their own information-management standards. The countries and regions where data is actually processed, and detailed contractual conditions, were not disclosed in this announcement.
The companies have also described policies for how the service will be provided; results of an independent security audit by a third party have not been published.
Early technical collaboration with semiconductor makers has already begun, but concrete measured figures such as the general availability date, pricing, design-time reduction, and PPA improvement rates have not been published.
The announcement also includes no case of a chip designed with GPT-Synopsys passing sign-off and going into mass production.
To judge the benefits of adoption, it will be necessary to look not only at how fast the AI generates proposed fixes but also at the total design time, including re-verification and engineer review.
Even if AI makes it possible to try many design candidates, the overall benefit to development could shrink if the computing resources needed for analysis grow accordingly or if redoing mistaken fixes takes time. The price of a service that bundles computing resources and EDA licenses will also shape adoption decisions.
What to watch for in future customer case studies is how much time and cost to sign-off can be cut under the same design conditions while maintaining or improving power, performance, and area.
If such figures are shown, the value of GPT-Synopsys can be assessed not merely at the level of "it can try many design candidates" but by how much it shortens and streamlines actual chip development.