Digital Asset
Intelligence
Theme

Search 3CBlock

News, exchanges, regulatory intelligence, reviews and Voices.

OpenAI Models Rewrite AMD and Jalapeño Chip Circuitry, Claiming 14% PPA Gains

Add us on Google

Summary: OpenAI reports its AI models achieved up to 14% improvements in power, performance, and area on AMD GPU IP and its own Jalapeño chip, compressing a year of design progress into five months while acknowledging self-improvement risks.

Crypto editorial illustration by 3CBlock
Illustration: 3CBlock

OpenAI announced on November 18 that its artificial intelligence models are now actively rewriting the circuitry of its own custom hardware and advanced third-party semiconductors. The company stated that this automated design process yielded up to 14% improvements in power, performance, and area (PPA) on intellectual property for AMD’s flagship GPUs, which OpenAI described as already heavily optimized.

Accelerated Design Cycles

According to a November 18 research post, OpenAI’s models compressed approximately one year of traditional engineering progress into less than five months during the AMD collaboration. Some of the circuit modifications generated by the AI are already in the manufacturing pipeline, while others have been archived in production files for future silicon iterations. OpenAI further noted that AMD confirmed single-digit percentage PPA gains across several key blocks of its next-generation flagship CPU.

The technical workflow involved the models rewriting SystemVerilog Register Transfer Level (RTL) code, which describes a chip’s logical circuits. After each rewrite, the system used Synopsys verification tools to ensure that the block’s behavior remained intact. This marks a significant expansion of OpenAI’s hardware design efforts, which began in early 2025 when the company directed its GPT-o3 mini model at the Jalapeño custom inference chip, co-developed with Broadcom. At that time, OpenAI reported that model-proposed changes reduced standard-cell area by about 13% in the chip’s vector unit and 10% in its matrix unit.

In separate evaluations involving 9,881 scored runs on PicoRV32 and Ibex open-source RISC-V processor cores, OpenAI found that its newest GPT-6.2 Astra model recorded the highest average scores among five tested generations. The company did not disclose a detailed breakdown of the three PPA measures for the AMD results, nor did it name the specific AMD product involved.

Safety Frameworks and Criticism

OpenAI is currently developing GPT-Synopsys, a specialized model built in partnership with Synopsys to operate electronic design automation (EDA) tools directly. However, the company acknowledges the inherent risks of this feedback loop. AI self-improvement, defined as letting models lead the development of new models, is flagged in OpenAI’s Preparedness Framework as one of three frontier capability areas monitored for severe harm risks, alongside biological/chemical threats and cybersecurity.

The framework commits OpenAI to halt development if a model hits a "Critical" self-improvement threshold until appropriate safeguards are in place. This approach has drawn scrutiny from external researchers. In a May 2026 analysis, LessWrong researcher Charbel-Raphaël argued that OpenAI’s self-improvement red line is fundamentally flawed. The analysis contends that the control mechanism "fires too late," relies on self-certification, and lacks external evaluators, noting that OpenAI has not clearly defined what constitutes mitigations meeting the critical standard.