OpenAI improves chip design with AI
OpenAI says its models reduced the area of key blocks in its Jalapeño chip and achieved improvements of up to 14% in AMD designs. They also outperformed in tests with RISC-V processors and are now working with Synopsys to integrate with professional design and verification tools.

OpenAI says its models are already optimizing real chips, from its custom Jalapeño processor to AMD designs. The improvements affect performance, power consumption, and the area occupied by the circuits, three factors that determine how much a chip costs and how quickly it works.
The company also published evaluations using open RISC-V processors and announced a collaboration with Synopsys to connect its models to the tools chip engineers use.
Improvements to the Jalapeño chip
OpenAI began using GPT-4o in early 2025 to help design Jalapeño, its first custom chip for running AI models. The system found changes in key parts of the processor, including the engine that manages parallel tasks and the unit that performs matrix operations.
In one improvement, a multiplier used approximately 15% less area and reduced its latency from two clock cycles to one. Another redesign, applied to an adder, also reduced area by around 15% without changing its two-cycle latency.
According to the team’s estimates, the changes proposed by the models reduced standard-cell area by approximately 13% in the parallel execution engine and 10% in the matrix unit. OpenAI says these optimizations were validated in silicon during Jalapeño’s bring-up, meaning the process of checking the physical chip after it was manufactured.
Results with AMD hardware
OpenAI also worked with AMD to optimize existing designs. In AMD’s main graphics unit, its models achieved improvements of up to 14% versus already optimized baselines in PPA, an indicator that combines power, performance, and area.
The model rewrote SystemVerilog RTL code, which describes how circuits work before they are manufactured. It then used Synopsys tools to verify that the changes preserved the correct behavior and to measure their effects on power consumption, speed, and size.
OpenAI compares the result with approximately one year of progress achieved in less than five months, although this comparison applies only within the project and under its evaluation conditions. AMD has already incorporated some changes into a design sent for manufacturing and saved others for future versions.
The collaboration also produced single-digit percentage improvements in important blocks of AMD’s next flagship CPU, according to the company. Work continues on graphics and central processors.
What the RISC-V tests show
To measure how its models are advancing, OpenAI tested them with two open 32-bit processors based on RISC-V: PicoRV32, designed for reprogrammable chips, and Ibex, used in the OpenTitan security chip.
In one test, each model received the original design and had to improve its performance. OpenAI checked that the processors continued to work correctly and discarded designs that failed or exceeded 16 times the original area. Performance was calculated using instructions executed per second on workloads the models had not seen.
With a comparable budget of approximately 112,000 tokens per run, GPT-6.2 Astra achieved the highest average score on both processors. GPT-6.1 Sol also outperformed GPT-5.6 Sol in both cases. The full evaluation included 9,881 scored runs.
In another test, the models did not receive an existing design to improve. They received only written requirements and connection templates. All five evaluated models performed better when creating a core from those instructions than when optimizing one that had already been built.
That result has an important explanation: starting from a specification allowed the model to omit features present in the original designs, such as performance counters or additional instructions. OpenAI therefore acknowledges that this test does not yet fully represent production chip design. It would need to require those features, impose tighter area limits, and check timing after physical synthesis.
What changes for you
Chip design usually requires months of specialized work. If models can find smaller, faster, or more efficient circuits, the effect could eventually reach the products you use: data centers with lower power consumption, cheaper AI services, and devices with more capacity in the same amount of space.
But the model does not manufacture the chip by itself. It proposes changes, writes them, and submits them to automated checks and engineering review. Manufacturing remains a physical, expensive, and slow process in which a single error can force the entire design to be repeated.
OpenAI plans to work with Synopsys on GPT-Synopsys, a model trained to use electronic design automation tools in the same way expert designers do. The next point to watch will be whether these improvements hold up as real-world constraints increase: mandatory functions, strict limits on area, power consumption, manufacturing schedules, and full silicon validation.

