AI Rewrites the Low-Level Code of Chip Design: OpenAI Model Achieves 14% PPA Gain on AMD Flagship GPU, Some Changes Already Sent for Tape-Out

OpenAI says its model rewrote SystemVerilog code for an AMD flagship GPU IP and achieved up to a 14% PPA improvement, with some changes already included in

On November 18, 2026, OpenAI disclosed a chip design achievement on its official blog: its model automatically rewrote SystemVerilog code describing circuit structures and achieved up to a 14% improvement in the combined power, performance, and area (PPA) metric for an AMD flagship GPU IP. OpenAI said this is equivalent to completing in under five months optimization progress that normally takes about a year, and that AMD has incorporated some model-generated changes into the latest chip design sent for manufacturing, while other changes have been written into production design files for future versions.

What the Specific Numbers Mean

PPA is a core optimization target in the chip design industry, and balancing all three dimensions at once is extremely difficult—typically, reducing power sacrifices performance, and shrinking area may raise the frequency ceiling. The 14% OpenAI claims is the upper bound of combined improvement; the official blog did not break down the respective contributions of power, speed, and area.

This figure is already exceptional in the industry. Modern top-tier GPU designs undergo several rounds of deep optimization before tape-out, and the so-called “heavily optimized baselines” leave extremely limited room for further improvement. Squeezing a double-digit PPA improvement out of that baseline, regardless of methodology, deserves serious attention.

A technical report on the United Nations AI research website unite.ai disclosed more granular figures: in the matrix unit, the multiplier redesigned by the model not only reduced standard-cell area by 15% but also cut latency from two clock cycles to one; the adder design likewise reduced area by 15% while maintaining two-cycle latency. These specific changes show that the model’s optimization is not a vague overall “compression” but a structural reworking of specific functional blocks.

Jalapeño: The History of AI Involvement in OpenAI’s In-House Inference Chip

In addition to the AMD GPU work, OpenAI’s official blog also disclosed design details of Jalapeño, its first in-house inference chip. The changes proposed by the model reduced standard-cell area by about 13% in the vector unit and about 10% in the matrix unit.

According to reports, Jalapeño took only nine months from initial design to tape-out, and AI assistance was one of the key factors in compressing the timeline. OpenAI initially used GPT-4o in Jalapeño’s design work, meaning the chip’s full lifecycle spans model iterations from GPT-4o to the GPT-6 series.

This forms a quantifiable positive feedback loop: the design efficiency of AI inference chips partly comes from AI models trained on the previous generation of AI hardware.

RISC-V Benchmark: Model-Generation Differences Behind 9,881 Scoring Runs

To evaluate different models’ chip design capabilities, OpenAI conducted systematic tests on two open-source RISC-V processor cores (PicoRV32 and Ibex), recording a total of 9,881 scoring runs. The tests were divided into two task types: RTL optimization (improving existing circuit code) and spec-to-RTL generation (directly generating circuit descriptions from design intent).

Under a budget constraint of about 112,000 tokens, GPT-6.2 Astra achieved the highest average score on both tasks. On the PicoRV32 RTL optimization task, GPT-6.2 Astra scored about twice as high as GPT-6 Astra. On the spec-to-RTL generation task, GPT-6.2 Astra averaged above 0.7 on PicoRV32 and about 0.75 on Ibex.

OpenAI proactively acknowledged an important limitation in the blog: RTL generated from specifications may omit some functions present in the original core (such as performance counters and additional instruction sets), so scores on the spec task cannot be directly equated with the quality of production designs.

Synopsys Collaboration: Striking at the Core of the EDA Toolchain

OpenAI also disclosed its collaboration with chip design tool giant Synopsys: the two are jointly training a model called GPT-Synopsys, with the goal of enabling AI to operate electronic design automation (EDA) tools like a seasoned chip engineer. The multi-year collaboration was officially announced on September 30, 2026.

EDA tools are the “compiler” of chip design—translating engineers’ design intent into manufacturable physical layouts. Synopsys is the undisputed leader in the global EDA market, and its toolchain is embedded in the design flows of nearly all mainstream chips. If GPT-Synopsys can effectively operate this toolchain, the coverage of AI-assisted design will expand from “optimizing existing circuit code” to “end-to-end participation in the entire flow from design to manufacturing.”

Data security in joint training was also explicitly addressed: OpenAI and Synopsys stated in the blog that “customers’ design data will not be used to train” the model, with encryption protection and configurable permission, audit, and retention controls.

Interests: OpenAI Holds a Potential Stake of About 10% in AMD

When assessing this 14% figure, the commercial context cannot be ignored.

According to reports by CNBC and other media in October 2025, OpenAI signed a multi-year chip supply agreement with AMD: OpenAI committed to deploying a total of 6 gigawatts (GW) of AMD GPU compute, while AMD issued OpenAI warrants to purchase up to 160 million AMD shares at an exercise price of just $0.01 per share. If OpenAI exercises them fully, it will obtain about 10% equity in AMD, with vesting milestones tied to deployment scale and AMD stock price targets (up to $600 per share).

This means OpenAI has a direct financial interest in AMD’s stock price rising. And an announcement that “AI significantly improved AMD’s chip design efficiency” has a potential positive impact on AMD’s stock price. AMD has incorporated some changes into actual tape-out, providing external validation.

This announcement on OpenAI’s official blog does not contain any direct statement or endorsement from AMD. In July 2026, AMD said the two sides “are jointly researching how AI can improve systems that run AI,” but AMD did not offer independent comment on the specific figures released on November 18.

What Is the Industry Significance?

For the past few years, discussions of AI-assisted chip design stayed at the tool level: AI helps verify, helps debug, and helps generate test cases. What would truly change the landscape is whether AI can directly generate RTL code that can be sent for tape-out—that is the lowest-level language through which design intent is translated into physical circuits, and one of the most technically formidable stages in chip engineering.

What OpenAI disclosed this time is that model-generated SystemVerilog changes have already gone into AMD’s tape-out version. Crossing this boundary is more substantively meaningful than any benchmark score.

From an industry-structure perspective, if this path proves viable, the most directly affected parties are not chip design companies like AMD or Nvidia—they are beneficiaries of this technology. The affected parties are the heavily manual optimization steps in traditional EDA flows, and mid-tier chip design service companies whose core competitiveness rests on “being able to tune better parameters.”

From OpenAI’s own strategic logic, this is a multi-win move: higher design efficiency for Jalapeño directly lowers in-house chip costs; the PPA improvements generated by the AMD collaboration financially reward OpenAI (through the AMD stock appreciation channel); and the Synopsys collaboration paves the way for promoting this capability across the industry. Progress at each link feeds back into the design of the next generation of models and chips.

Independent Assessment

This announcement is more credible than typical AI capability promotion for three reasons: the numbers are specific enough (down to the area and latency of the matrix-unit multiplier); there is external validation (the changes went into AMD’s manufacturing version); and OpenAI proactively disclosed methodological limitations (spec-to-RTL tasks may omit functions).

But the real question is not whether the numbers are true, but how far this path can go. Is completing a year’s worth of optimization progress in five months because manual work missed optimization headroom in the existing baseline, or because AI is systematically exploring solution spaces beyond human designers’ cognitive blind spots? If it is the latter, this speed advantage will accelerate further as AI involvement deepens, rather than mean-revert.